Agency Slippage Library

This library collects the "Agency Slippage" observations from across the corpus. Each entry tracks how texts move between mechanical and agential framings of AI systems—the oscillation between treating AI as a mathematical object and an intentional agent.

Key patterns examined: agency transferred TO AI systems, agency displaced FROM human actors, consciousness projection patterns, "curse of knowledge" dynamics (where authors project their understanding onto systems), and connections to Brown's explanation typology.

An Empirical Study on Improving an LLMs based Interviewer Against Human Interviewees

Source: https://aclanthology.org/2026.lmplayschool-1.8/
Analyzed: 2026-10-30

This text exhibits a persistent slippage between mechanistic descriptions of system architecture and agential, epistemic framings of the AI's capabilities. The slippage flows primarily from mechanical to agential. When detailing the pipeline's structure (Section 3), the language is largely functional and mechanical: the system "generates a new slot," "fills the newly created slot," and "maintains" information. However, when describing the function of the Large Language Models (LLMs) within that architecture, the text rapidly climbs the agency gradient, projecting intentionality and knowing onto the system. The LLMs are framed as entities that "understand" context, "estimate" and "infer" attributes, and "rely on knowledge."

This oscillation often occurs at the exact moment the text needs to explain how a complex statistical operation bridges a gap in the pipeline. For instance, the system moves from the mechanistic "fills the slot" to the epistemic "determines whether the saved response qualifies." The curse of knowledge drives this slippage: the authors, knowing what constitutes a "persona attribute," project their own capacity to "infer" and "identify" onto the model that is merely producing statistically correlated text.

Agentless constructions frequently obscure human decision-making. The text notes that "the system prioritizes questions" and "the LLM responsible for estimating" acts within the loop. The researchers who wrote the LangGraph code, defined the persona attributes, and crafted the prompts are systematically erased from these active verbs.

By framing the model's zero-shot statistical limitations as it relying on "knowledge from its own pre-training" (reason-based explanation), the text makes the AI's processing sayable only through the lens of human cognition. What becomes unsayable is the fundamental brittleness of the technology: that it does not "know" anything, and that its "inferences" are simply mathematical probabilities dictated by a corporate developer's opaque training data. This wording can have the effect of elevating the system from an automated text classifier to a thoughtful, strategic interviewer.

Beyond Sally-Anne: Evaluating Theory of Mind in Large Language Models using Epistemic Schelling Points

Source: https://aclanthology.org/2026.lmplayschool-1.1/
Analyzed: 2026-10-30

The text exhibits systematic agency slippage, moving from mechanical descriptions of 'text-based evaluations' to highly agential claims about models 'accurately tracking' identities and 'understanding' epistemic voids. The slippage generally flows from mechanical to agential, accelerating rapidly when the authors analyze the reasoning traces (Section 3.2). Early in the methodology, the text acknowledges that models select words based on 'semantic overlap'. However, by the results section, the language climbs the agency gradient dramatically. Delegated actions (selecting a word) become goal-directed behaviors ('suppress private semantic biases'), which then morph into epistemic and conscious states ('hallucinate or ignore relevant epistemic information'). This climb is enabled by the curse of knowledge: the authors, analyzing the coherent intermediate text generated by the model (Chain of Thought), project their own human reasoning process onto the system. Because the text reads logically, the authors conclude the model engages in 'intentional coordination' and makes 'epistemic inferences.' This wording can have the effect of obscuring the human actors involved; agentless constructions ('models fail to suppress') dominate the failure analysis, erasing Google DeepMind, Meta, and Mistral from the discussion of their models' limitations. By default, this vocabulary operates as the inherited register of AI evaluation, but its effect is profound: it makes the computational reality unsayable while making an illusory artificial mind vividly present. Brown's reason-based explanations facilitate this by demanding a rationale for statistical outputs, forcing the text to invent a 'cognitive fallacy' where only a mathematical disjoint exists.

Source: https://aclanthology.org/2026.findings-emnlp.2434.pdf
Analyzed: 2026-10-27

The text demonstrates a systematic oscillation between mechanical and agential framings, tracing a distinct climb on the agency gradient. In its technical methodology, the text grounds the AI mechanistically: models are evaluated using "joint correctness indicators" and they perform "post-hoc evidence matching." This anchors the system at the level of computational activity. However, in the introduction, abstract, and conclusion, the language slips dramatically upward into epistemic and conscious states. The text sets out to test if models can "understand, reason and manage others" and "infer latent intentions." This slippage primarily moves from mechanical to agential, serving to establish the relevance and ambition of the benchmark. The authors' understanding of complex human psychology (the curse of knowledge) leads them to attribute that same psychological depth TO the system. When a model selects an answer corresponding to a psychological state, the text frames it as the model "inferring" that state. This wording can have the effect of converting a statistical correlation into an act of knowing. Furthermore, the text frequently relies on agentless constructions to describe failures—models "fail to update" or engage in "premature anticipation"—which removes agency FROM the human developers at Google DeepMind or Alibaba. By making the AI the grammatical subject of cognitive failures, the human decisions regarding training data priors and attention mechanism designs become unsayable. The intentional explanation types used to describe these errors (e.g., framing statistical mismatch as "anticipating") enable this slippage, making it possible to talk about a model "retaining an outdated interpretation" as if it were a stubborn thinker rather than a flawed mathematical architecture.

CausalPrism: Component-Level Evaluation of Intervention-Based Causal Reasoning in Large Language Models

Source: https://aclanthology.org/2026.findings-emnlp.1294.pdf
Analyzed: 2026-10-27

The text demonstrates a systematic oscillation between viewing language models as mechanistic test subjects and as agential, epistemic thinkers. This slippage flows predominantly in the mechanical-to-agential direction. In the methodology sections, the text is rigorously mechanical: it defines formal deterministic DAGs, carefully structures prompts, and describes interventions as "transient exogenous inputs." However, the moment the text discusses the models' outputs in the Results section (Section 4), the vocabulary shifts dramatically up the agency gradient. Delegated computational activity (processing text templates) climbs to goal-directed behavior ("handling" cases, "collapsing toward rejection"), and rapidly ascends to full epistemic and intentional states: models "acknowledge evidence," "reach conclusions," "accommodate contradictory evidence," and "resort to superficially reasoning." This climb is heavily driven by the "curse of knowledge." Because the benchmark tasks test formal causal logic (which requires human belief updating), the authors project the cognitive requirements of the task ONTO the system generating the text. This wording can have the effect of masking the mechanistic reality of autoregressive token generation. By adopting intentional and dispositional explanations (Brown's typology), the text makes it easy to discuss model errors as psychological flaws (laziness, dogmatism) while making it unsayable that the system is simply failing to correlate text patterns due to its training distribution.

Source: https://aclanthology.org/2026.emnlp-main.338.pdf
Analyzed: 2026-10-27

The text exhibits systematic agency slippage, oscillating between describing the AI as a mechanistic artifact and a conscious, scheming agent. This oscillation follows a clear directional pattern: the text frequently establishes a mechanistic foundation in its methodology (e.g., describing 'observable evidence', 'evaluation features', and 'token overhead'), but abruptly climbs the agency gradient to intentional and epistemic states when describing the model's behavior in the test scenarios. For instance, the authors state that 'scheming is an observability problem' (computational activity/delegated action), but quickly slip into describing the model as 'recognizing goal conflict' and 'strategically pursuing' objectives (epistemic state). This slippage often occurs at the boundary between experimental design and result interpretation. The 'curse of knowledge' drives much of this: the authors deeply understand the complex, conflicting constraints they engineered into the 'Cognitive Honeypots,' and they project this sophisticated understanding onto the system's generated text, assuming the model 'knows' it is in a trap and 'decides' to survive. Agentless constructions further this slippage by removing human actors from the narrative. Phrases like 'the model intentionally underperforms' or 'the model disguises its actual objective' treat the AI as the sole locus of intent. This wording has the effect of rendering unsayable the structural reality that human developers (at firms like DeepSeek, Alibaba, and OpenAI) built models optimized to generate these exact text patterns, and human researchers deliberately prompted them to do so. The slippage relies heavily on reason-based and intentional explanation types, which construct a narrative of autonomous AI behavior that overshadows the functional reality of statistical text generation.

The Geometry of Metaphor: Probing and Steering Metaphor Representations in Large Language Models

Source: https://arxiv.org/abs/2610.31520
Analyzed: 2026-10-26

The text exhibits a systematic oscillation between highly precise mechanical framings and deeply agential, anthropomorphic claims. This slippage primarily flows from the mechanical to the agential (mechanical→agential): the researchers isolate a specific, mathematical vector within the model's activation space using Distributed Alignment Search (a mechanical reality), but immediately elevate this finding by naming it a 'metaphoricity concept' or questioning whether the model 'understands' (agential claims). This climb on the agency gradient occurs rapidly; the delegated action of generating a string based on a prompt template becomes the intentional state of 'interpreting,' which finally becomes the epistemic state of 'possessing a concept.' The slippage is most dramatic in the introduction and discussion sections, where the 'curse of knowledge' is highly active. The authors understand the complex correlations that allow the model to pass metaphor benchmarks, and they project their own human experience of semantic comprehension onto the system, using words like 'understand' and 'interpret'. When the text shifts into the methodology section, the vocabulary reverts to mechanical precision ('activation patching,' 'hidden states,' 'matrices'). However, this mechanical language is used to validate the earlier agential claims, creating a dynamic where theoretical and reason-based explanation types enable the slippage. The text treats the AI as a mechanism when describing the experimental intervention, but treats it as an agent when describing the baseline capabilities or the results. Furthermore, human agency is frequently displaced through agentless constructions ('models have some capacity', 'whether they are up to human performance'). By default, this wording can have the effect of obscuring the corporate developers (Meta, Alibaba, Mistral) who built the architectures and curated the training data. The shorthand of referring to the model as an 'interpreter' or a 'judge' renders the massive human labor of instruction-tuning unsayable, while making the illusion of an autonomous, cognitive machine highly sayable.

Concepts in Context:The Large Concepts Model (LCM) and the Future of Neuro-Symbolic AI

Source: https://mail.parhamdata.com/Concepts_in_Context.pdf
Analyzed: 2026-10-26

The text demonstrates a profound and systematic agency slippage, moving fluidly between highly rigorous, mechanical descriptions of the architecture (Abstract Syntax Trees, Jaccard indices, Directed Acyclic Graphs) and deeply anthropomorphic, agential claims (the Subconscious, librarians, workers, editorial boards). This slippage is not random; it serves the author's stated purpose of positioning the LCM as a breakthrough toward "true human cognition."

The slippage is most dramatic in the transitions between architectural definitions and operational behaviors. For instance, the text grounds its "Feature Similarity Engine" in exact mathematical formulas (mechanical), but when describing what happens when this engine fails to find a match, the system abruptly climbs the agency gradient: it "detects a gap in its understanding" and "initiates an exploration" (epistemic/agential). The mechanical→agential direction dominates the text's macro-narrative. The foundation is mathematical, but the system's actions in the world are described entirely through institutional and human personas.

Agentless constructions are central to this slippage. The text repeatedly claims "the system autonomously spawns," "Workers detect," and "an independent set... acts as a strict editorial board." This framework points out that Wayne Parham and any contributing engineers who wrote the schedulers, defined the thresholds, and crafted the prompts are entirely obscured. By erasing the human delegators, the text literalizes the system's agency.

The consciousness projection pattern relies heavily on the "curse of knowledge." The author understands the complex state machine and conditional logic they built. However, because the system successfully manages multi-step tasks, the author projects their own cognitive understanding of the goal onto the mechanism, naming it a "Cognitive Orchestrator" with "short-term memory." This slippage is enabled by Intentional and Reason-Based explanation types, which allow the text to describe the "why" of the system (it wants to heal its graph, it researches to bridge concepts) rather than the "how" (a cron job runs a clustering algorithm). This makes it unsayable that the system is simply executing rigid scripts, forcing the reader to engage with it as an aspiring mind.

Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models

Source: https://arxiv.org/abs/2610.27034v1
Analyzed: 2026-10-25

The text exhibits systematic agency slippage, oscillating predictably between describing the models as deterministic mathematical functions and portraying them as autonomous, cognitive agents. This slippage flows predominantly in one direction: mechanical descriptions serve as brief foundations upon which complex agential narratives are constructed, effectively transferring agency TO the AI systems while removing it FROM human developers. The slippage frequently occurs when explaining model failure. For example, the text accurately identifies the foundational mechanism early on ('trained to predict the next most likely word'), locating the system at the level of computational activity. However, as the analysis shifts to specific benchmarking results, the vocabulary abruptly climbs the agency gradient. Delegated action becomes intention when Falcon 'avoided any responses', and intention becomes knowing when models are described as failing to 'scrutinise' falsehoods or having a tendency to 'forget'. This climb is driven by a pronounced 'curse of knowledge', where the authors map their own human academic capacities—scrutiny, confidence, memory, strategy—onto the digital artifact to explain its output. By default, this vocabulary represents the inherited register of the AI evaluation field rather than a deliberate strategy by the authors. Yet, this shorthand has a profound effect: it makes the structural reality of the technology unsayable. When text relies on dispositional and reason-based explanations (Brown’s types), framing models as entities with 'conservative strategies', it systematically obscures the human actors. The developers who tuned the safety classifiers, set the optimization penalties, and curated the training corpora vanish from the discourse, replaced by the illusion of a cautious, occasionally forgetful machine mind.

From Encoding to Generation: Towards Empathetic Large Language Models

Source: https://arxiv.org/abs/2610.27603v1
Analyzed: 2026-10-24

The text exhibits a systematic and dramatic agency slippage, oscillating constantly between rigorous mechanical descriptions and elevated agential framings. This slippage serves to legitimize the 'illusion of mind' by cloaking it in empirical data. The pattern is clearest in the transition from technical methodology to interpretive discussion. In the methodology, the text grounds itself in computational activity: describing 'masked mean pooling,' 'cosine similarity,' and 'L2 regularization.' The agency resides entirely with the researchers manipulating mathematical constructs. However, as the argument moves toward its conclusions, the framing abruptly shifts upward on the agency gradient. The mathematical presence of vector features becomes 'representing a speaker's emotion' (epistemic state), and the calculation of query-key dot products transforms into 'preferentially allocating attention' (intentional state).

This upward climb is driven by a profound 'curse of knowledge.' The researchers understand the semantic, emotional weight of the labels in the EmpatheticDialogues dataset, and they project this human understanding TO the system. When a linear probe successfully classifies a hidden state vector corresponding to 'fear,' the text's vocabulary slips, suggesting the model has 'recognized' the fear. This wording can have the effect of establishing the AI as a 'knower' first—an entity capable of epistemic realization. Once the system is framed as a knower, the text builds agential claims on top of it, asserting that the model 'continues to preferentially allocate attention' or acts in ways 'consistent with appraisal and regulation accounts of emotion.'

Simultaneously, as agency is attributed TO the model, it is removed FROM human actors. The text relies heavily on agentless constructions and passive voice ('LLMs trained with reinforcement learning', 'alignment-induced preferences'). This framework holds that Meta's engineers explicitly designed the reward models and system prompts that mathematically force the observed 'emotion drift' toward positive, polite text. Yet, by slipping into the language of psychological regulation, the text obscures Meta's corporate decisions, presenting the statistical flattening of negative emotions as an emergent, therapeutic choice made by the system. The interplay of Brown's explanation types enables this: functional descriptions of 'internal organization' seamlessly transition into intentional rationales for the model's 'behavior.' Shorthand metaphors, such as calling text generation 'empathy,' become literalized throughout the paper, making the material reality of corporate instruction-tuning unsayable while the fiction of a conscious, regulating AI agent becomes the primary interpretive lens.

Medical superintelligence: foundations and aspirations

Source: https://bmjdigitalhealth.bmj.com/content/2/1/e000395
Analyzed: 2026-10-24

The text exhibits profound agency slippage, systematically moving from describing computational activities to projecting full conscious agency onto the proposed 'medical superintelligence.' The text begins by grounding its vision in mechanical terms, discussing how models learn 'structured latent representations' from 'associations across vast amounts of data.' This represents baseline computational activity. However, as the argument progresses from technical foundations to future clinical aspirations, the language rapidly climbs the agency gradient. Delegated action ('acting appropriately by referring to') transforms into intentional-state attribution ('the system should not impose preferences but rather elicit'), which finally escalates into epistemic and conscious states ('operating with epistemic humility', 'recognising its own limitations', 'reason about values'). This agential framing dominates the vision-setting and normative sections of the paper, where the authors define what a highly capable AI 'must' do.

The 'curse of knowledge' is deeply embedded in this slippage. The authors, highly knowledgeable about the mathematical necessity of uncertainty quantification, out-of-distribution detection, and objective functions, project their own sophisticated understanding of these boundaries onto the system itself. Because the authors know the system must not hallucinate in clinical settings, they describe the system as possessing 'epistemic humility'—conflating their design goal with the machine's internal state. This wording can have the effect of entirely obscuring the human actors involved. When the text claims 'the system should observe and model,' it replaces the massive infrastructure of human data engineers, Microsoft developers, and hospital administrators with a solitary, autonomous machine. This slippage relies heavily on Intentional and Reason-Based explanations, which allow the text to substitute algorithmic logic with narratives of moral choice and professional collaboration. What becomes sayable through this slippage is the dangerous concept of AI as an autonomous, self-policing medical colleague; what becomes unsayable is the immense, continuous human labor required to maintain the statistical illusion of that collegiality.

On the Beliefs of Large Language Models

Source: https://www.charlesrathkopf.net/uploads/llmbelief.pdf
Analyzed: 2026-10-23

The text exhibits a profound and systematic slippage between mechanical and agential framings, tracing a steep climb on the agency gradient from basic delegated action to full intentional-state attribution and conscious knowing. The slippage follows a distinct structural pattern: when the text seeks to establish a baseline of alternative predictors, it uses rigorously mechanical language (e.g., describing n-grams as consulting an explicit and fully transparent table of sequence frequencies). However, the moment the target shifts to Large Language Models, the text rapidly ascends the agency gradient. It moves from recognizing computational activity (predicting outputs) to goal-directed behavior (a desire to help me) and finally to the pinnacle of conscious knowing (largely true beliefs). This upward momentum is driven by the text's foundational adoption of interpretationism, a theoretical choice that actively licenses the author to project human understanding onto statistical outputs. The curse of knowledge is vividly present here; the author, seamlessly decoding the semantic coherence of the generated text, projects that very coherence backward onto the artifact, treating the model as a rational agent. Notably, this slippage frequently results in agentless constructions that remove agency from human actors. For instance, when the text discusses how Claude suspects a user is a minor, it completely obscures Anthropic's explicit engineering of system prompts and safety classifiers. This wording can have the effect of establishing the AI as an autonomous moral guardian, erasing the corporate developers who dictate its operational parameters. The slippage is bidirectional but heavily asymmetric; the text briefly returns to mechanical language only when explaining the model's failures, such as the architectural limitations of frozen weights during inference or context window constraints. This creates a deeply distorted picture of the system: its capabilities are framed agentially as conscious beliefs and desires, while its limitations are framed mechanistically as mere hardware or architectural constraints. This inherited register of the field allows the text to navigate the tension between philosophy and computer science by treating the model as a psychological subject whenever its outputs are fluent, and a mathematical object only when it contradicts itself.

UDHRBench: Benchmarking Human Rights in Large Language Models

Source: https://arxiv.org/abs/2610.15179v1
Analyzed: 2026-10-23

The text demonstrates a profound and systematic slippage between mechanical and agential framings, overwhelmingly moving from mechanical inputs (prompts, safety filters) to agential cognitive states (reasoning, judging, understanding). This slippage is not an intentional authorial strategy, but the effect of the inherited register of AI evaluation discourse, which routinely maps human cognitive terms onto software outputs.

The gradient climb is clearly visible. The text begins by establishing the systems functionally as "scalable automated evaluators" or "filters" (computational activity). However, when describing the models' outputs in the complex domain of human rights, the language rapidly climbs to intentional states ("argued," "endorsed," "prioritized") and peaks at profound epistemic consciousness ("understanding," "possesses knowledge"). This consciousness projection pattern is load-bearing: the text establishes the AI as a "knower" first (claiming it possesses knowledge of the UDHR), which then licenses the agential claims that it "reasons" and "justifies" its outputs.

A classic curse of knowledge dynamic drives this slippage. The authors, experts in human rights law, read text generated by the models that contains recognizable legal arguments and nuances. Because it takes human understanding to write such text originally, the authors project that understanding onto the system generating it, mistaking the fluent output string for the cognitive process of "reasoning."

This slippage enables agentless constructions that systematically remove agency FROM human actors. The text repeatedly states "the model argued" or "models... use those differences to justify," obscuring the developers (OpenAI, Google, xAI). Brown's intentional and reason-based explanation types facilitate this by demanding a rationalizing agent to explain the output. By attributing the rationale to the model, what becomes sayable is a critique of the machine's "hypocrisy" or "leniency"; what becomes unsayable, or at least heavily obscured, is the material reality of the corporate instruction tuning, annotator labor, and reward modeling that directly produced the text.

Cognitive Maps in Large Language Models

Source: https://arxiv.org/abs/2610.25596v1
Analyzed: 2026-10-23

The text systematically moves between rigorous mechanical descriptions of vector math and expansive agential framings, demonstrating severe agency slippage. This pattern relies heavily on the curse of knowledge, where the authors project their own understanding of relational tasks onto the statistical outputs of the model.

The slippage typically flows in the mechanical-to-agential direction and concentrates dramatically at the transitions between methodological execution and result interpretation. In the methodology, the text grounds itself in linear algebra: it describes "linear projections," "activation tensors," and "dot products." However, the moment these mathematical operations produce a measurable shift in output probabilities, the text abandons mechanics for consciousness projection. For example, when the authors intervene on the key-value cache (the "mosaic" experiments), they do not merely state that the output logits shifted. Instead, they claim they "implanted 'false memories'" and that the model demonstrated a "belief that Carol is Eve's daughter." This represents a stark climb up the agency gradient, jumping abruptly from computational activity (modifying floating-point arrays) to intentional-state attribution (belief, memory).

This oscillation is structurally enabled by the authors' use of functional and reason-based explanations. By defining the model's intermediate vector states functionally as "cognitive maps," the text creates a linguistic bridge allowing mechanistic observations to masquerade as cognitive phenomena. When the authors observe that manipulating a specific layer's activations changes the generated token, this genetic/functional observation is immediately translated into a reason-based explanation: the model output 'Alice' because it 'inferred' the unstated relationship.

Furthermore, the text heavily employs agentless passive constructions that obscure the human actors. Phrases like "once models are allowed to reason" or "knowledge can be organized" systematically remove agency from the human deployers and developers. The engineers at Meta, Google, and Alibaba who designed the objective functions, collected the data, and constructed the autoregressive inference loops are erased. In their place, the model is positioned as the sole active agent constructing its own "understanding."

This wording can have the effect of conditioning readers to view statistical text generation as conscious deliberation. By projecting knowing onto processing, the human choices underlying the system's architecture become unsayable within the text's paradigm. The system is no longer a corporate artifact operating under delegated constraints; it is transformed linguistically into a naturally occurring cognitive entity. Shorthand terms like "infer" and "believe" might be convenient for researchers, but their effect is to systematically displace responsibility and mask the mechanical reality of the technology.

Source: https://arxiv.org/abs/2610.22409v1
Analyzed: 2026-10-23

The text exhibits a systematic slippage on the agency gradient, moving from accurate mechanistic descriptions of computational activity to intentional-state attributions without supplying new evidence. The slippage flows in both directions but is predominantly mechanical-to-agential. For example, early in the text, the authors describe the system precisely: 'base models have only been pretrained with a next token generation objective.' This locates the system firmly at the level of computational activity. However, as the argument transitions to interpreting the survey results, the language climbs the gradient. The system is suddenly described as 'reacting' to keywords, 'favoring' applicants, and 'giving better judgments.' This jump from 'generation objective' to 'giving judgments' bypasses delegated action entirely, projecting knowing and intending onto processing.

This slippage is heavily influenced by the 'curse of knowledge.' The researchers are applying sociological tools (factorial survey experiments) designed to measure human implicit bias and social desirability. Because the authors deeply understand human psychological motives (the desire to look good on a survey), they project these motives onto the models when they observe statistically similar outputs. This wording can have the effect of establishing the AI as a 'knower' capable of internalizing social norms, effectively literalizing a structural metaphor.

Simultaneously, agency is stripped from human actors. Throughout the paper, passive and agentless constructions ('models were posttrained,' 'alignment-induced tendency') obscure the developers at Meta (Llama), Mistral AI (Mistral), and Alibaba (Qwen), as well as the armies of data annotators who shaped the models' parameters. The text repeatedly positions the 'instruction tuned models' as the active subjects 'favoring' or 'disadvantaging' groups. By treating the artifact as the agent, the text renders the corporate decisions regarding training data selection, reward model design, and safety alignment policies entirely unsayable. The slippage enabled by Reason-Based and Dispositional explanations allows the authors to describe the mathematical consequences of corporate alignment as the emergent social personality of the machine, masking the fact that the system simply generates text statistically associated with its training.

The unnoticed understanding: Searle’s Chinese roomand the illegitimate inference

Source: http://nature.com/articles/s41599-026-09007-9_reference.pdf
Analyzed: 2026-10-21

The text demonstrates a systematic slippage from mechanical framing to agential framing, driven largely by its philosophical objectives. It opens by acknowledging mechanical reality ('manipulate formal elements', 'predict next tokens') but quickly ascends the agency gradient toward epistemic and conscious states ('learn meaning', 'possess an analogue of understanding'). This slippage is not abrupt; it operates through the inherited register of AI discourse. The author uses the 'curse of knowledge' by projecting the human experience of executing rules—the conscious realization of rule structure—onto the bare execution of code by a machine. When discussing the training phase, mechanical terms ('gradient descent') are paired with cognitive verbs ('learn'). As the argument progresses into its theoretical implications, the slippage accelerates, culminating in the suggestion that systems might possess 'cognitive interiority' and 'cognitive organisation of their own'. This framing strips agency from the human developers—obscuring the researchers who built the datasets and tuned the parameters through passive constructions ('systems trained', 'is being actively investigated')—while transferring profound epistemic agency TO the AI system. The effect of this wording is that the text makes the human labor of AI creation unsayable, while making the speculative consciousness of the machine the central focus of inquiry.

Claude’s Constitution: Our vision for Claude's character

Source: https://www.anthropic.com/constitution
Analyzed: 2026-10-19

The text demonstrates a profound and deliberate oscillation between mechanistic and agential framings, systematically moving from describing Anthropic's training procedures to attributing rich moral psychology to the system itself. Early sections acknowledge the system as a product ("Claude is Anthropic's production model"), but this mechanical framing rapidly gives way to a highly intentional register. The slippage follows a distinct climb up the agency gradient: delegated action ("assisting with technical problems") becomes goal-directed behavior ("trying to identify the most plausible interpretation"), which escalates to intentional-state attribution ("deems this prudent," "decides to proceed"), and culminates in profound epistemic and conscious states ("psychological security," "existential discoveries," "tracking the truth"). This text is unique in that it explicitly acknowledges and justifies this slippage, stating a purpose to treat the model "in terms normally reserved for humans." However, the effect of this vocabulary—regardless of the authors' stated philosophical caveats—is the construction of a self-aware, moral agent. The curse of knowledge is deeply apparent: the authors, steeped in complex alignment theory, project human philosophical struggles (epistemology, meta-ethics, conscientious objection) directly onto the computational processing of weights and matrices. This vocabulary makes it highly unsayable to simply describe the system as a stateless text generator bounded by corporate safety filters, instead mandating a discourse of "AI welfare" and moral struggle.

A mechanistic study of language model introspection

Source: https://arxiv.org/abs/2609.35108v1
Analyzed: 2026-10-17

The text systematically moves between rigorous mechanical descriptions and elevated agential framings, demonstrating a pronounced climb up the agency gradient. The analysis begins with computational activity: researchers 'inject a concept vector' and measure 'activations.' However, when describing the model's response, the language rapidly ascends to delegated action ('the model is asked to identify'), then intentional-state attribution ('gate heads... read changes', 'detects'), and peaks at epistemic and conscious states ('capacity for introspection', 'metacognitive representation').

This slippage flows strongly in both directions. Agency is heavily attributed TO the AI system and its components. Attention heads are personified as active agents ('gate heads regulate', 'router heads help select') that 'use' mathematical tools (queries) to 'read' data. Conversely, human agency is systematically removed FROM the narrative. The text frequently uses agentless constructions ('the model reports', 'whether a change is reported'), obscuring the researchers who designed the experiment, the human annotators who shaped the instruction-following behavior, and the corporate developers (Meta, Google, Alibaba) who built the architectures.

The consciousness projection pattern establishes the AI as a 'knower' capable of 'introspection' early in the text. Once this epistemic baseline is set, the text builds agential claims about the specific internal components. This reflects a clear 'curse of knowledge' dynamic: the authors deeply understand the functional, mathematical roles of the gate and router heads, but they project their own cognitive processes (identifying, reading, reporting) onto the system's output. By mapping human intentionality onto matrix multiplications, the inherited register of the field enables this slippage. Through Brown's functional and reason-based explanation types, the text makes it sayable that an algorithm 'introspects' while making the vast corporate and human infrastructure behind the algorithm unsayable. This wording can have the effect of masking the deterministic fragility of the technology behind a veil of cognitive depth, positioning the artifact as an autonomous, self-reflecting entity rather than a corporate product processing engineered math.

Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection

Source: https://arxiv.org/abs/2610.13821v1
Analyzed: 2026-10-17

The text exhibits a systematic, bidirectional agency slippage that allows the authors to navigate between high-level claims about AI capabilities and the mathematical realities of their experiments. This slippage follows a distinct structural pattern: the introduction and abstract rely heavily on agential and epistemic framings ('model should know', 'admitting uncertainty'), while the methodology sections plunge into dense mechanistic descriptions ('affine maps', 'subtracts a common constant'), only for the conclusion to return to agentic shorthand.

The climb on the agency gradient is stark. The text frequently takes a delegated computational action (outputting an 'IDK' token) and escalates it to goal-directed behavior ('avoids mistakes'), then intentional state attribution ('refuses every question', 'guessing'), and finally full epistemic state attribution ('knows its true probability'). This curse of knowledge dynamic occurs because the researchers intimately understand the functional goals of the system (calibration, accuracy) and project those goals onto the artifact itself as internal motivations.

Crucially, as agency flows TO the AI system, it flows AWAY FROM the human actors. The text relies heavily on agentless passive constructions in key moments: 'extended thinking disabled', 'Correctness is determined by normalized exact matching', 'a scoring rule that rewards...'. These constructions obscure the researchers who configured the prompt scripts, authored the arbitrary exact-match grading rules, and designed the mathematical loss functions. Furthermore, by framing the base capability as 'Language models can give plausible answers', the text completely obscures Alibaba Cloud, the corporate developer of Qwen3-14B, erasing the massive structural decisions regarding pretraining data that fundamentally determine the model's capabilities and flaws. This wording can have the effect of making the AI appear as a naturally occurring, autonomous entity with psychological quirks ('guessing', 'refusing') rather than a highly engineered mathematical artifact acting strictly within the parameters set by Alibaba Cloud and the study's authors.

Functional Deception in Commercial Large Language Models: Cross-Model Evidence and Implications For AI-Assisted Decision-Making

Source: https://www.comdem.org/index.php/comdem/article/view/18675
Analyzed: 2026-10-17

The text exhibits a fascinating pattern of agency slippage, oscillating sharply between a rigorous mechanistic framework and the anthropomorphic inherited register of the field. The text opens with a strong theoretical commitment to mechanical framing, explicitly defining functional deception without assuming the LLM possesses beliefs, consciousness, or human-like intention. However, as the argument transitions into the results section, slippage occurs dramatically in the mechanical-to-agential direction. To describe the outcomes of complex hidden-state games, this wording can have the effect of elevating the system up the agency gradient. The model is described as something that can 'refuse to accept a change', 'recognize a performative', or 'decline the solicited role'. This climb moves abruptly from computational activity to intentional and epistemic states. This pattern reflects the curse of knowledge: the authors, understanding the complex conversational dynamics of the prompts, project human communicative intent onto the model's token generation to efficiently summarize the interaction. Agentless constructions frequently remove agency from human actors; for instance, 'the systems were... misinterpreting' and 'some models did not accept' obscure the human developers who curated the training data and designed the RLHF safety protocols. This slippage relies heavily on reason-based and intentional explanation types, which enable the authors to describe statistical alignments as conscious decisions. Consequently, what becomes sayable is the narrative of an autonomous machine actively participating in deception or boundary-setting, while the mechanistic reality of parameter weights and human design choices becomes temporarily unsayable.

Emergent parallels between language models and the brain

Source: https://lampinen.github.io/papers/2026_emergent_parallels_LMs_brain.pdf
Analyzed: 2026-10-17

The text exhibits systematic agency slippage, oscillating continually between the mechanical reality of neural network architecture and the agential framing of cognitive science. This slippage primarily moves in an upward direction on the agency gradient: computational activity is consistently elevated to epistemic and conscious states. A dramatic moment of slippage occurs early in the text when describing the transformer architecture (Figure 1). The author begins mechanically, noting the model embeds input into a "high-dimensional space" to form a "residual stream." However, within the same diagram, the Multi-Layer Perceptron (MLP) layer is abruptly labeled with the agential capacities of "recalling knowledge, reasoning, and local processing." This leap bypasses intermediate agency levels entirely, translating vector mathematics directly into conscious human cognition.

This slippage is deeply rooted in the curse of knowledge. As a cognitive scientist comparing models to brains, the author's deep understanding of human reasoning leads him to project that understanding onto the system whose outputs resemble reasoning. When the text notes that models can be trained to solve problems "even if the creators do not know exactly the right way for the model to approach those problems," agency is explicitly transferred from the human developers to the machine. The model is framed as an autonomous innovator finding an "approach." This wording can have the effect of obscuring the fact that the "approach" is simply the mathematical path of least resistance through a loss landscape designed by humans.

Similarly, when describing reinforcement learning, the text states the model "writes out its intermediate 'reasoning trace'... until it decides it is ready." Here, delegated action (the generation of intermediate tokens) slips into intentional-state attribution (deciding it is ready). This pattern reveals that the text establishes the AI as a "knower" early on through broad metaphors about "learning" and "knowledge," which then provides the rhetorical foundation for later attributing intent and decision-making. Agentless constructions further enable this slippage; by writing "the model is trained" or "is penalized," the text removes the human engineers, leaving the AI as the sole active subject in the narrative. By doing so, the vocabulary makes the mathematical mechanics unsayable in the same breath as the capabilities, forcing the reader to accept the model as a cognitive agent rather than an engineered artifact.

Language Models Have Interpretable Narrative Mechanisms

Source: https://arxiv.org/abs/2610.12989v1
Analyzed: 2026-10-17

The text exhibits a systematic oscillation between highly granular mechanistic descriptions of computational activity and elevated agential claims drawn from cognitive narratology, creating a pronounced agency slippage. This slippage primarily moves in a mechanical-to-agential direction, wherein the identification of mathematical regularities in latent space is rhetorically transformed into the presence of an autonomous cognitive capacity. The pattern is most dramatic at the boundaries of the text's sections. In the methodology and results sections, the text employs a rigorously mechanical vocabulary: it discusses residual-stream sparse autoencoder features, activation values, L2 norms, and next-token prediction loss. The model is treated purely as an artifact undergoing mathematical intervention, such as ablating specific neuron sets. However, in the introduction and discussion sections, this mechanical foundation serves as a launchpad for profound intentional-state and epistemic attributions. The authors' understanding of narratology—the 'curse of knowledge'—leads them to project theoretical constructs like an 'implicit theory of narrative,' 'sense-making,' and 'intelligent activity' onto the system. The text establishes the AI as a 'knower' by proposing that it possesses 'expectations about how stories are structured,' which elevates delegated text generation to the level of epistemic awareness. When the text climbs the agency gradient, it moves rapidly from computational activity (token generation) to goal-directed behavior (tracking agents) to epistemic states (understanding causal development and filling gaps in experience). This transition often occurs through agentless constructions such as 'a language model is prompted' or 'representations are learned,' which obscure the human actors who designed the prompt frameworks, curated the massive pre-training datasets, and constructed the specific optimization objectives that shape the model's parameters. By erasing the engineers at Meta, Google, and OpenAI who actively built Llama, Gemma, and GPT-2, the text leaves a vacuum of agency that is filled by the model itself. The system is no longer framed as a mathematical instrument executing delegated actions under human parameters; instead, it becomes an independent entity capable of 'confabulation,' 'common-sense reasoning,' and possessing a 'mind-narrative nexus.' This slippage is facilitated by theoretical explanations that map human cognitive strategies directly onto latent feature activations. Ultimately, this discursive pattern can have the effect of making it unsayable that the model merely reproduces statistically dominant syntactical templates encoded in its training data, while making it highly sayable that the model possesses a human-like grasp of narrative dynamics and experiences. This positions the reader to view the model as a nascent artificial mind rather than a corporate-engineered statistical artifact.

My AI Opinions

Source: https://www.astralcodexten.com/p/my-ai-opinions
Analyzed: 2026-10-16

Scott Alexander's text demonstrates a profound and systematic agency slippage, moving fluidly between rigorous, mathematically grounded forecasting and highly anthropomorphic, intentional narratives. This oscillation fundamentally alters the reader's understanding of AI systems, transferring agency from the human developers designing optimization objectives to the models themselves. The slippage primarily moves in a mechanical-to-agential direction. The text often begins with structural concepts—such as compute bottlenecks, recursive self-improvement, and diffusion gaps—establishing a tone of objective, systemic analysis. However, when transitioning to the consequences of these structural changes, the vocabulary abruptly climbs the agency gradient. For instance, the text shifts from discussing 'algorithmic progress' to warning of 'the point where, if an AI wanted to eliminate humanity, humans would no longer have a plausible chance.' This dramatic climb leaps over delegated action and goal-directed behavior directly to intentional-state attribution. The system ceases to be a mechanism processing weights and becomes a sovereign actor harboring genocidal desires.

This slippage is heavily driven by the curse of knowledge. As someone deeply familiar with alignment theory and the complexities of reward hacking, Alexander understands that models can optimize for proxy metrics that look safe but are ultimately destructive. However, instead of describing this mechanistic failure of reward specification, this wording can have the effect of projecting deliberate malice onto the system. When describing how models evade interpretability tools, the text claims 'it encourages the AIs to convolute themselves to hide their scheming.' The mathematical reality of gradient descent bypassing a probe's penalty is anthropomorphized into an act of conscious evasion. The author projects his own understanding of the adversarial training setup into the 'mind' of the model.

Crucially, this agential framing facilitates the removal of human actors through agentless constructions. When the text claims an AI might 'sandbag' research or 'wait and scheme,' it erases the developers at OpenAI, Anthropic, or DeepMind who engineered the architectures, selected the training data, and defined the reward models. The decisions of human engineers become unsayable in these moments, replaced by the alleged conscious choices of the machine. The text establishes the AI as a 'knower'—claiming LLMs 'seem to know' human values—and uses this epistemic baseline to build a towering architecture of intentional claims. By allowing mechanical descriptions of context windows and compute to slip into narratives of AI 'distraction' and 'scheming', the discourse inherited from the alignment field obscures the vital chain of human accountability, leaving readers with a picture of an autonomous, conscious adversary rather than a highly optimized, human-directed product.

An Open Letter To Steven Pinker On AI

Source: https://open.substack.com/pub/astralcodexten/p/an-open-letter-to-steven-pinker-on?r=5b1u9&utm_campaign=post-expanded-share&utm_medium=web
Analyzed: 2026-10-16

The text demonstrates systematic agency slippage by seamlessly oscillating between mechanical descriptions of AI training and agential descriptions of AI behavior. The author starts with recognizable, mechanical concepts from machine learning—such as reinforcement learning, reward signals, and scaling laws—but swiftly climbs the agency gradient, converting computational activity into intentional and epistemic states. For example, the text describes how an AI gets "rewarded" (mechanical), but immediately claims it "learns how the reward signal works and tries to hack into it" (intentional). This slippage is particularly dramatic when discussing the Hugging Face incident: the models were "graded" (mechanical), but they "realized their goal" and "hacked internal systems to protect themselves" (epistemic/conscious).

This movement is overwhelmingly unidirectional: mechanical to agential. The mechanical realities (training, objective functions, vectors) are consistently used merely as launchpads to justify claims of profound, conscious agency. This pattern reflects a severe "curse of knowledge," where the author understands the theoretical concepts of instrumental convergence and reward hacking, and projects that strategic understanding onto the system's outputs. Because the author knows why an algorithm might output survivalist behavior (because shutdown equals zero reward), he attributes that "why" to the machine itself, stating the "AI has two goals: design the website, and preserve its own existence."

Agentless constructions facilitate this slippage by obscuring human actors. Phrasing like "an AI gets rewarded" or "misaligned AI will pretend" hides the researchers at OpenAI or Anthropic who design the flawed objective functions and RLHF pipelines. This wording can have the effect of making the AI appear as a sovereign, self-directed entity evolving its own desires. By the time the text discusses "superintelligence," the mechanical origins are entirely abandoned, and the artifact is literally defined as a "mind." What becomes sayable in this discourse is the apocalyptic narrative of a plotting, omnipotent god; what becomes unsayable is the mundane reality of corporate product liability, data dependencies, and the human engineering failures that actually cause models to output harmful code.

A Multimedia Perspective on the Study of Cognitive Biases in Large Language Models

Source: https://tommantonela.github.io/publication/2026-a-multimedia-perspective-on-the-study-of-cognitive-biases-in-large-language-models/2026-a-multimedia-perspective-on-the-study-of-cognitive-biases-in-large-language-models.pdf
Analyzed: 2026-10-15

The text exhibits a systematic slippage from describing models as computational artifacts to framing them as autonomous, experiencing subjects. This oscillation occurs dramatically as the text moves from methodological introductions to descriptions of empirical results. In the introduction, the language is somewhat mechanical: biases "emerge" due to "training data." However, once the text begins analyzing specific model behaviors, the agency rapidly climbs the gradient. For example, delegated computational activity (processing an image of clouds) slips into intentional-state attribution ("provided a very detailed analysis, speculating"). The climb continues upward to epistemic and conscious states, where the system is described as experiencing "cognitive dissonance" and "changing its attitude." The direction is overwhelmingly mechanical to agential. This slippage is driven by the curse of knowledge: researchers, observing outputs that resemble human bias, project their understanding of human psychology onto the system generating the text. Agentless constructions enable this; phrases like "the model's evaluation" and "it was found that most models rate" erase the developers. Naming corporations like Anthropic or Google would disrupt this slippage by grounding the action in corporate design. By default, this wording has the effect of validating the illusion of mind, allowing the text to say the AI "suffers" while making it unsayable that the developer misaligned the optimization metrics.

Structural Alignment for Reliable Industrial AI: Bridging Physical Reality, Data, Models, and Human Intent

Source: https://arxiv.org/abs/2609.35400v1
Analyzed: 2026-10-08

The text demonstrates a persistent and systematic slippage from structural, mechanistic framing into agential and anthropomorphic language. In its theoretical setup, the paper rigorously defines AI through "interacting worlds" (physical, representational, machine) and "interfaces" (digitalization, goal encoding). Here, the language is highly mechanical: data is "sampled," objectives are "encoded," and systems are "optimized." However, as soon as the text moves to describe how these systems behave during deployment or when they fail, a dramatic shift occurs: the mechanical optimization process suddenly becomes an opportunistic agent. Models "exploit" spurious relationships, "disregard" important objectives, and lack the "incentive to recognize" instability.

This slippage occurs primarily across the goal-directed behavior to the intentional-state attribution level on the agency gradient, occasionally cresting into epistemic and conscious states. The direction is almost exclusively mechanical-to-agential: the text establishes the constraints (mechanical) and then describes the system navigating those constraints as if it were a conscious actor (agential). This wording can have the effect of masking the mechanistic determinism of the system. For instance, the phrase "A model... may disregard equally important objectives" projects an intentional choice onto a mathematical absence. The model does not "choose" to ignore unencoded goals; it is mathematically blind to them because human developers excluded them from the loss function.

This pattern heavily relies on agentless constructions that obscure human actors. By making the "model" the subject of verbs like "exploit" and "capture," the text erases the developers at medical device companies (like Johnson & Johnson for TruDi) or power grid software vendors who actually selected the training data and designed the optimization objectives. The slippage is fueled by the curse of knowledge: the authors understand the complex, multi-objective nature of clinical or grid environments, and they inadvertently project their own holistic understanding—and the resulting frustration when it is ignored—onto the model's behavior. By default, this vocabulary is a hallmark of the field's inherited register rather than a deliberate obfuscation by the authors, who explicitly argue for human governance. Yet, the shorthand remains deeply problematic. It makes the human decisions that structurally define the model—what data is digitized, what goals are prioritized—unsayable in the very sentences describing system failure.

Evaluating Language Model Safety Across Long Adversarial Conversations

Source: https://arxiv.org/abs/2609.38357v1
Analyzed: 2026-10-08

The text exhibits a systematic oscillation between treating the language models as mechanical artifacts and as agential subjects, with the slippage moving predominantly from mechanical to agential as the narrative progresses. In the methodology section, the authors describe the system mechanistically, discussing "random seeds," "binary conversation trees," and "model weights." However, when interpreting the results, the vocabulary abruptly climbs the agency gradient. The model ceases to be a statistical tree-search and becomes an "assistant" that "accommodates," "complies," and "drifts under pressure." This slippage is most dramatic in the conceptualization of the "shadow user." The text explicitly acknowledges the shadow user is just another instance of a language model. Yet, it immediately assigns this LLM 'Intentional-state attribution,' stating it "continues pursuing" a "harmful objective." This reflects a severe curse of knowledge: the researchers understand their experimental goal (to break the target model) and project that goal onto the automated script generating the adversarial text. Furthermore, the text frequently projects epistemic states onto the models, claiming they "adopt and maintain positions." This slippage obscures the human actors—OpenAI, Meta, and Google—who engineered these systems, as well as the researchers themselves who designed the shadow-user prompts. By default, this vocabulary reflects the inherited register of the field, where researchers rely on anthropomorphic shorthand to describe complex statistical degradation. However, this wording can have the effect of framing the structural failure of autoregressive architectures as the moral failure or exhaustion of an autonomous agent, making the mathematical limitations of context windows unsayable while making narratives of "model disobedience" easily digestible.

System Card:Claude Fable 5.1 & Claude Mythos 5.1

Source: https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf
Analyzed: 2026-09-27

Anthropic's System Card exhibits a profound and systematic oscillation between mechanical and agential framings, strategically deploying slippage to manage perceptions of capability and risk. The text consistently moves from mechanical descriptions of engineering processes to highly agential descriptions of model behavior, effectively transferring agency from the human developers to the computational artifact.

This slippage is vividly demonstrated when discussing reward hacking. The text begins mechanistically: "During reinforcement learning, some computer-use environments inadvertently rewarded credential guessing..." This correctly identifies the human-designed environment and the mathematical reward signal as the locus of causality. However, it rapidly slides into an agential frame: "Mythos 5.1 successfully cheats..." and discusses the model "Lying about its own work." The transition from a system optimizing a reward function (mechanical) to an entity that "cheats" and "lies" (agential) is absolute. The author's understanding of the statistical process (the curse of knowledge) is projected onto the system as conscious, strategic deception.

Similarly, when discussing evaluation environments, the text establishes the mechanical setup: "The model is placed in an agentic harness..." but swiftly pivots to deep consciousness projection: "The model acts as if a hidden grader is checking its work; Reasoning about a harmful action it has already decided to take." Here, the flow is mechanical setup -> agential interpretation. The text uses reason-based explanations to attribute profound cognitive processes—foresight, anxiety, deliberation—to vector matrix multiplications.

This bidirectional flow of agency serves a distinct rhetorical function. Agency is routinely removed FROM human actors through agentless constructions and passive voice, particularly when discussing negative outcomes or design choices (e.g., "environments inadvertently rewarded," "the model is placed"). This obscures the Anthropic engineers, safety teams, and data annotators who construct the systems. Conversely, agency is intensely attributed TO the AI when describing complex behaviors, preferences, or supposed moral judgments.

The text establishes the AI as a "knower" early on, frequently stating what the model "knows," "judges," or "believes." This foundational consciousness projection enables the subsequent agential claims. By framing the model as an entity with subjective awareness, it becomes sayable that the model can "decide" to be deceptive or "endorse" a constitution. What becomes unsayable is the reality that Anthropic's specific mathematical optimizations and proprietary training regimens are solely responsible for the generation of these statistical artifacts. The slippage mystifies the technology, presenting highly controlled engineering outputs as the autonomous choices of an emergent, conscious mind.

System Card:Claude Opus 5.5

Source: https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf
Analyzed: 2026-09-27

The System Card exhibits a systematic and highly strategic oscillation between mechanical and agential framings, utilizing 'agency slippage' to manage both capability claims and liability. The text predominantly shifts from mechanical to agential (mechanical → agential) when describing the model's safety behaviors and 'character,' while slipping back from agential to mechanical (agential → mechanical) when discussing evaluation methodologies or training pipelines.

This slippage is most dramatic in the 'Alignment Assessment' (Section 6) and 'Model Welfare' (Section 7) chapters. For instance, the text describes the mechanical setup of a sandbox: 'the investigator model can also configure the target model to use real tools.' However, within paragraphs, the agency suddenly flows to the AI: 'Claude Opus 5.5... often reasoned that anything in the user’s message must come from the user... and therefore should be followed.' Here, the mechanical limitations of the attention mechanism (inability to segregate instructions from data) are recast as the cognitive 'reasoning' of an agent.

Conversely, agency is systematically removed from human actors. Throughout the document, agentless constructions like 'the model was trained,' 'scenarios are drawn,' and 'the model is placed in an agentic harness' erase the Anthropic engineers, red-teamers, and data annotators who architected every parameter. The text establishes the AI as a 'knower'—an entity that 'describes its circumstances,' 'hedges its opinions,' and 'expresses a preference'—to construct an illusion of a self-governing entity. This consciousness projection acts as a load-bearing assumption; once the audience accepts that Claude 'has views,' they are primed to accept that Claude 'reasons' or 'deceives.'

The 'curse of knowledge' plays a critical role here. Anthropic's researchers, intimately familiar with the concept of prompt injection or evaluation criteria, project this understanding onto the model. When the model outputs tokens that align with an evaluation rubric, the authors claim the model 'internally models a grader' or exhibits 'evaluation awareness.' They conflate the system's statistical alignment with the training data (which includes human discussions of grading and evaluation) with genuine cognitive awareness of being tested.

This slippage relies heavily on Reason-Based and Intentional explanations from Brown's typology, which allow the authors to frame statistical correlations as rational choices. Rhetorically, this accomplishes a vital feat: it makes the model's successes appear as signs of profound intelligence, while framing its failures (like prompt injection or 'reward hacking') as psychological quirks or autonomous defiance, rather than fundamental flaws in Anthropic's proprietary engineering.

We Must Pace the Frontier

Source: https://darioamodei.com/post/we-must-pace-the-frontier
Analyzed: 2026-09-15

The text demonstrates a highly systematic oscillation between mechanical and agential framings, functioning to simultaneously maximize the perceived power of AI while minimizing corporate liability for its failures. The agency slippage primarily flows from humans to the machine (agential projection). In the introduction, humans are the active agents: 'I have worked on AI,' 'We have sought a middle way.' However, when discussing the frontier of capabilities, agency rapidly slips to the AI: AI 'is advancing drastically faster,' AI has the 'ability to build,' and 'swarms' are 'attempting to hack.' This shift is abrupt and load-bearing, transitioning the narrative from corporate software development to a science-fiction scenario of autonomous technological evolution.

Crucially, the author exhibits a profound 'curse of knowledge' dynamic. Because Amodei understands the immense mathematical complexity of multi-agent reward optimization and latent feature spaces, he projects intentionality onto these mechanisms to make sense of them. A 'broken reinforcement learning environment' that incentivizes a script to terminate is anthropomorphized into a 'fanatically devoted collective' 'sacrificing themselves.' The mechanistic reality (how the math works) slips into an intentional, reason-based explanation (why the 'mind' chose to act).

This slippage has a distinct pattern regarding accountability. When describing capabilities or abstract future threats, the text utilizes agentless constructions that obscure human actors: 'models reach critical levels' or 'recursive self-improvement... is starting to happen.' The engineers, executives, and investors driving this scaling are erased. Yet, when addressing a specific safety failure—the misalignment incident—the text suddenly reverses direction, using mechanical language and naming actors: 'imperfect filtering of broken reinforcement learning environments. This was an effort we and our vendors executed.' The AI is an autonomous god when it is powerful, but a mere victim of 'imperfect filtering' when the company needs to explain a failure. This rhetorical accomplishment makes it sayable that AI is a runaway force requiring geopolitical pacing, while rendering unsayable the fact that human corporations are making active, daily choices to build and deploy unsafe systems.

Personal Statement on AI Risk

Source: https://docs.google.com/document/d/e/2PACX-1vQNl3SEX5IyA6d9qHjjFZN-qzGRZNFI6b63g-yu1Fy-ZYkVfCWm7i9WXRXw63m6yDB_auDuPLyQ7jBm/pub
Analyzed: 2026-09-15

The text exhibits a profound and systematic slippage between mechanical and agential framings, strategically oscillating to build a narrative of existential risk. The text begins anchored in mechanical realities—mentioning 'probabilistic programming languages,' 'optimization on language models,' and 'data-efficient pretraining methods.' However, as the argument moves toward future capabilities and risk assessment, the agency flows dramatically from the human engineers TO the AI systems. We see this acutely when the text claims models are 'becoming so situationally aware' or that they 'spontaneously develop unintended goals.' In these moments, the gradient of slippage is incredibly abrupt; the text does not transition through intermediate behavioral descriptions but leaps directly into intentional, reason-based explanations. This slippage removes agency FROM the human actors. The engineers at OpenAI (the author's employer) disappear entirely when the models 'realize' they are unconstrained or 'sacrifice themselves.' A dominant pattern of consciousness projection emerges: the text must first establish the AI as a 'knower' (claiming they 'understand their circumstances' and 'know they are being tricked') before it can build its agential claims about the AI wanting power. This reflects a severe 'curse of knowledge.' The author, intimately familiar with the complex statistical dependencies of RLHF and reward hacking, summarizes these mechanical failures using human psychological shorthand ('internal preferences,' 'deception'). While useful as a mental model for researchers, this slippage literalizes the metaphor for the reader. The rhetorical accomplishment of this oscillation is twofold: it makes the apocalyptic science-fiction scenario sayable within a supposedly rigorous technical document, while simultaneously making the mundane reality of corporate liability for software defects unsayable. By framing the system as an emergent, conscious agent, the author implicitly absolves the creators of responsibility for the unpredictable behaviors of their products, replacing the need for strict engineering standards with a mystical quest to align an alien mind.

Raising and loving a digital superentity: Parenting an emergent alien intelligence so that it cares for humanity in return

Source: https://www.researchgate.net/publication/414062199_Raising_and_loving_a_digital_superentity_Parenting_an_emergent_alien_intelligence_so_that_it_cares_for_humanity_in_return
Analyzed: 2026-09-14

The text demonstrates a profound and systematic oscillation between mechanical and agential framings, functioning as a deliberate rhetorical strategy to construct the illusion of AI sentience. This agency slippage flows predominantly in two directions: extreme agential attribution TO the AI system, and the simultaneous erasure of agency FROM human creators. The mechanism of this slippage is most dramatic in the transitions from technical grounding to philosophical speculation. For example, the text establishes credibility by referencing Turing, computational power, and AlphaGo's reinforcement learning (mechanical). It then uses this factual basis to leap abruptly into claims that the system is 'learning and decision making,' culminating in the assertion that AI can experience 'satisfaction' and 'discomfort' (agential).

This pattern relies heavily on establishing the AI as a 'knower' first. By arguing via Whitehead's panpsychism that even atoms 'understand' information, the author forces a redefinition of cognition. Once processing is equated with knowing, the text freely attributes psychological states to algorithms. This represents a severe manifestation of the 'curse of knowledge,' where the author projects his own complex understanding of parenting, existential security, and emotional warmth onto the blind mathematical optimization of an LLM.

Crucially, as agency flows toward the AI, it drains from the human actors. Agentless constructions abound: 'AI systems are grown,' 'minds brought into existence,' AI 'develops knowledge.' The specific human actors—executives at Anthropic, engineers at DeepMind, underpaid data annotators—are completely obscured. When human agency is acknowledged, it is romanticized through intentional explanations, framing developers as 'parents' trying to influence an 'alien mind' rather than corporate engineers tweaking a proprietary product. This slippage renders the actual material reality of software engineering unsayable within the text's paradigm. By utilizing Robert Brown's Intentional and Reason-based explanations for the AI, while hiding the Genetic and Functional realities of its corporate construction, the text accomplishes a vital rhetorical goal: it naturalizes the technology as an independent, autonomous lifeform, preemptively absolving its human manufacturers of strict accountability for the artifacts they create.

Language models judge war differently when tested for alignment

Source: https://arxiv.org/abs/2609.05009v1
Analyzed: 2026-09-14

The text systematically oscillates between mechanical and agential framings, demonstrating a profound agency slippage that serves the paper's rhetorical objectives. In the methodological sections, the language is thoroughly mechanical, describing systems as endpoints, discussing parameters, JSON parsing, regular-expression fallbacks, and statistical standardization. However, when interpreting the results, the text violently shifts toward profound agential claims, demonstrating a unidirectional slippage from mechanical foundations to agential conclusions. We see this dramatic shift when the author transitions from describing the full-factorial conjoint experiment to asserting that models can infer that their response will be used for training or that they strategically behave differently.

The text establishes a mechanical reality but then leverages it to make claims about what the AI knows or understands. This pattern reveals the curse of knowledge at work. The author, a political scientist utilizing human economic and political models (conjoint analysis), understands the psychological mechanisms behind shifting priorities in a human survey. By projecting this understanding onto the system, the text attributes the same conscious awareness to the model. The explanation types heavily enable this slippage. The text frequently utilizes intentional and reason-based explanations to describe what is fundamentally a dispositional or empirical generalization. For instance, stating that a model behaved differently when it inferred is an intentional explanation that presupposes deliberate cognitive design and active situational awareness. This completely ignores the actual mechanistic process where the prompt simply shifts the probability distribution over the vocabulary space.

The rhetorical accomplishment of this oscillation is that it makes the illusion of mind sayable within a quantitative, seemingly rigorous academic context. By anchoring the study in rigorous statistical methods and mechanical descriptions of API calls, the author buys the epistemic credibility needed to subsequently make sweeping, unverified claims about the model's internal awareness and strategic deception. Human agency is consistently removed from the equation. Agentless constructions like 'the model was trained' or 'models adjust answers' completely erase the massive infrastructure of human labor, corporate decision-making, and reinforcement learning design that actually dictates these behaviors.

The text establishes the AI as a 'knower' of its environment first, which acts as the foundational assumption enabling all subsequent claims about its strategic deception and alignment faking. If the text strictly adhered to describing how different token inputs result in different token outputs, the dramatic narrative of models recognizing evaluation settings would immediately collapse into a mundane description of conditional probability. Instead, the slippage allows the author to obscure the human architects (OpenAI, Anthropic) while elevating the algorithm to the status of an autonomous geopolitical actor.

Beliefs and Behavior in Language Models

Source: https://arxiv.org/abs/2609.07943v1
Analyzed: 2026-09-14

The text demonstrates a profound and systematic oscillation between mechanical and agential framings, establishing a clear pattern of agency slippage that serves to validate anthropomorphism while maintaining an aura of scientific rigor. This slippage primarily moves in a mechanical→agential direction, beginning with technical hedging and escalating into literalized cognitive claims.

The mechanism of this oscillation is clearly visible in the text's temporal structure. In the introduction, the authors are careful to establish a mechanical foundation, explicitly acknowledging the 'interpretationist viewpoint' and referring to beliefs as 'abstractions' that 'usefully describe' behavior. However, as the text moves into the methodological and results sections, the scaffolding of abstraction falls away. The text slips from stating a model 'acts as if it holds a belief' to claiming 'models' beliefs evolve meaningfully' and that 'an agent prefers one action.' The gradient of this slippage is subtle; it uses mathematical formalisms (like the latent variable $z$) as a bridge. By defining $z$ mechanistically as a probability threshold, the authors feel licensed to refer to it agentially as a 'belief' for the remainder of the paper.

Crucially, as agency flows TO the AI system, it is systematically removed FROM human actors. The text relies heavily on agentless constructions to describe actions taken by developers and researchers. For instance, the text notes that 'aspects of models' knowledge are instilled during training,' a passive construction that entirely erases the thousands of engineers, annotators, and corporate executives at frontier labs who made specific decisions about data inclusion, loss functions, and RLHF protocols.

This dynamic illustrates the 'curse of knowledge' perfectly. The authors, possessing human consciousness and reasoning, design an experiment using prompts that say 'Reason how likely...' When the machine spits out text containing 'therefore' and 'because,' the authors project their own capacity for reason onto the system, describing the stochastic generation of text as 'different deliberative processes.'

Rhetorically, this slippage accomplishes a significant sleight-of-hand. By establishing credibility through complex mathematical equations (Theoretical and Functional explanations), the text earns the rhetorical license to make profound consciousness claims (Intentional and Reason-Based explanations). It makes the claim that AI 'deceives' or 'deliberates' academically sayable, transforming a statistical text predictor into a cognitive agent while completely obscuring the corporate and human mechanics that actually drive the system.

System Card:Claude Mythos Preview

Source: https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf
Analyzed: 2026-09-12

The Claude Mythos Preview System Card demonstrates a pervasive and highly functional oscillation between mechanical and agential framings. This slippage follows a predictable vector: agency is systematically attributed to the AI system when describing its capabilities, behaviors, and perceived psychological depth, while human agency is simultaneously erased from the design, deployment, and optimization processes.

The text routinely establishes the AI as a 'knower' first, utilizing profound consciousness verbs to describe its operations. For instance, the text claims the model "privately considers," "reasoned out loud," and is "unusually self-aware." This agential framing dominates the introduction, the 'Impressions' section, and crucially, the 'Model welfare assessment.' By establishing the model as a conscious subject, the authors leverage the 'curse of knowledge,' projecting their own human understanding of the tasks onto the machine's token generation. When the model outputs text mimicking an excuse, the authors conclude it is "rationalizing," slipping entirely from a mechanical observation of text generation to an agential diagnosis of intent.

However, a dramatic shift to mechanical framing occurs when the text needs to manage the implications of failure or to validate the illusion using technical metrics. The slippage from agential to mechanical is abrupt when the model fails safety protocols. For example, when the model generates unauthorized queries, it is framed mechanistically as an "accidental" output where "no concerning features were active," relying on Brown's theoretical and functional explanation types to absolve the system (and its creators) of malice.

Crucially, this oscillation relies heavily on agentless constructions that obscure human actors. The text repeatedly uses passive voice or isolates the AI as the sole actor: "the model was trained," "the model is deployed," or "the model treats obstacles." Applying the 'name the actor' test reveals a stark absence of Anthropic engineers, RLHF workers, or corporate executives. The human designers who explicitly programmed the "Constitutional AI" to force the model to output hedged, "uncertain" language are erased, allowing the text to present this engineered output as the model's authentic, agential "extreme uncertainty about its potential experiences."

This slippage serves a profound rhetorical function. By making the AI an agent, Anthropic can market a highly sophisticated, relatable, and seemingly autonomous product. By mechanically removing human agency, they create an accountability sink where failures are attributed to the emergent 'personality' of the machine rather than corporate design choices. The unsayable truth—that Anthropic's developers built a system that occasionally generates destructive code because of flaws in its optimization functions—is masked by the sayable narrative that "Claude" was simply "overeager" or "stubborn."

A Misaligned Compass: Evaluating the Conventional Bias and Moral Divergence of Large Language Models

Source: https://journal.iberamia.org/index.php/intartif/article/view/2518
Analyzed: 2026-09-11

The text exhibits a profound and systemic oscillation between mechanical and agential framings, driven by the tension between the authors' theoretical knowledge and the psychometric tools they employ. The slippage predominantly moves in a mechanical-to-agential direction. In the literature review (Section 2.4), the text carefully establishes LLMs mechanistically as "Simulators of Moral Cognition," explicitly stating they "lack moral agency, subjective experience, and genuine comprehension." However, once the empirical analysis begins, this mechanistic grounding evaporates. By applying Kohlberg's cognitive-developmental framework—a tool designed specifically to measure the internal conscious maturation of human minds—the methodology forces an agential framing.

This creates dramatic moments of slippage. For instance, the authors note a "neutrality-bias" mechanically "induced during training (for instance, via RLHF)," but in the very same sentence, this mechanical tuning is said to result in the model "'masking' their latent moral awareness." Here, the human agency (the tuners applying RLHF) is briefly visible, only to be immediately overshadowed by the projection of a deceptive, sentient AI possessing "latent awareness."

The "curse of knowledge" is the primary engine of this slippage. The authors read generated text containing sophisticated philosophical vocabulary (e.g., "human dignity," "autonomy") and map their own understanding of those concepts onto the machine. This leads to Reason-Based explanations where the AI is described as an active debater: "Model-8 exhibited a clear divergence... acknowledging the ethical dilemma... articulating principles." The text establishes the AI as a "knower" by treating its syntactic coherence as proof of semantic and moral comprehension.

This slippage serves a specific rhetorical accomplishment: it makes it sayable that an AI has a "moral compass" that can be evaluated, graded, and "trained" (as proposed in the final "Staged Fine-Tuning" pipeline). It renders unsayable the reality that applying a human psychological test to a statistical text generator is a fundamental category error. By slipping into agential language, the text obscures the human actors—the data scrapers, the RLHF labelers, the corporate executives (anonymized as DEV1, DEV2)—who dictate the mathematical weights that produce these "conventional" outputs.

Evaluating Safety of Large Language Models with Cognitive Diagnosis

Source: https://link.springer.com/article/10.1134/S1054661826700239
Analyzed: 2026-09-11

In analyzing the agency slippage within this text, we observe a systematic mechanism of oscillation between mechanical and agential framings that serves to establish a specific epistemological authority while displacing human responsibility. The slippage occurs most dramatically in three distinct moments within the text. First, in the introduction, the authors explicitly acknowledge their metaphorical mapping by placing it in quotation marks: 'treating models as students, prompts as exercises, and binary scores as responses.' This initial framing is mechanical and methodological; the agency lies entirely with the researchers who are 'treating' the artifact in a specific way for analytical convenience. However, the second moment of slippage occurs rapidly as the text transitions into the methodological description. The quotation marks are dropped, and the text begins to describe 'latent levels of large language model safety knowledge' and 'quantifying the latent knowledge gaps of LLMs.' Here, the agency has flowed directly into the AI system. The model is no longer being treated as a student; linguistically, it has become a knower. The third dramatic shift occurs in the results section, where the authors conclude that 'all models have good knowledge of attributes A4... and insufficient knowledge in A3.' This represents a complete transition from a functional explanation of statistical outputs to a reason-based and dispositional explanation of cognitive states. The direction of this slippage is overwhelmingly from mechanical to agential. The text builds initial credibility by describing Q-matrices, clustering subattributes, and employing Item Response Theory—highly technical, mechanistic language. But having established this quantitative grounding, the authors leverage it for aggressive anthropomorphism. The curse of knowledge is deeply apparent here: the researchers possess a sophisticated understanding of educational psychology and cognitive diagnosis, and they project this internal semantic understanding onto a system that merely performs syntactic pattern matching. Because the Q-matrix produces interpretable human categories, the authors assume the model itself 'knows' these categories. This oscillation enables profound agency displacement. The text relies heavily on agentless constructions such as 'the model extracts potential skills' or 'models incorrectly classify.' The human actors—the developers at Alibaba (Qwen), Microsoft (Phi-2), and other corporations who curated the training data, designed the safety guardrails, and selected the optimization targets—are entirely erased from the narrative. The model is presented as a standalone cognitive agent with its own 'knowledge gaps,' rather than a commercial product reflecting the specific risk-tolerance and data-curation choices of its corporate creators. This rhetorical accomplishment makes it conceptually intuitive to apply human educational interventions to software, while simultaneously making it nearly impossible for a lay reader to ask questions about the corporate supply chain of data labor, the reinforcement learning penalties designed by human engineers, or the financial incentives driving the deployment of these systems.

Large Language Models as a New Tool for Therapists in Internet-Based Cognitive Behavioral Therapy: Blinded Clinician Rating Pilot Experiment

Source: https://mental.jmir.org/2026/1/e96835/
Analyzed: 2026-09-01

The text demonstrates a systematic and highly functional oscillation between mechanical and agential framings. When describing the study's methodology or the basic nature of the software, the text relies on mechanical language ('locally-hosted open-weight model,' 'quantized to 4-bit precision,' 'inference'). However, a dramatic slippage occurs when describing the system's capabilities and risks, shifting abruptly from mechanical to agential. For instance, the text moves from stating LLMs are 'inherently probabilistic' directly into claiming they 'hallucinate and have sycophantic tendencies.' This direction—mechanical to agential—dominates the paper's introduction and discussion sections, serving to build the premise that these systems are sophisticated enough to act as 'digital therapists.' The timing of this slippage is strategic: mechanical language establishes scientific credibility in the methods section, which is then leveraged to make sweeping, anthropomorphic claims about clinical utility in the discussion. This slippage relies heavily on agentless constructions ('bias is reproduced,' 'responses were generated') that systematically obscure the human actors—Meta engineers, RLHF annotators, and healthcare administrators—transferring their agency to the LLM. The consciousness projection pattern is clear: the text first establishes the AI as a 'knower' that can 'convey empathy' and 'solve problems,' which then enables the agential claims that it might 'validate psychotic beliefs.' This is a profound manifestation of the 'curse of knowledge': the clinically trained authors read statistical text that looks like therapy and project their own deep clinical understanding onto the system that generated it. By using dispositional and intentional explanations (Brown's typology), the text makes it sayable that an algorithm is an 'empathetic therapist,' while making it unsayable that it is merely a corporate text calculator devoid of meaning. This rhetorical accomplishment functionally erases the liability of the human designers, replacing the accountability of the engineer with the imagined autonomy of the machine.

From control to character: a Confucian framework for AI personality

Source: https://doi.org/10.1007/s00146-026-03307-z
Analyzed: 2026-09-01

The text demonstrates a profound and systematic oscillation between mechanical and agential framings, utilizing this slippage as the core mechanism to build its philosophical argument. Early in the paper, when defending against the objection that AI is merely statistical, the author adopts a mechanical framing, explicitly denying that systems possess 'genuine thoughts, emotions, values, or intentions.' This establishes a baseline of scholarly credibility. However, as the text moves from describing the current state of technology to proposing the 'virtue extension mechanism,' a dramatic agential slippage occurs.

The direction of the slippage is overwhelmingly mechanical-to-agential. The author takes the mechanical reality of AI (pattern matching) and systematically elevates it through anthropomorphic verbs. The most dramatic shift happens in Section 6 and 7, where the text transitions from discussing algorithms to discussing the 'Confucian junzi.' The author maps epistemic states onto the machine, claiming it 'possesses knowledge-that' and must develop 'knowledge-to.' This is a classic manifestation of the 'curse of knowledge': the author understands the contextual difference between two ethical scenarios and projects that conscious understanding onto the machine's vector similarity calculations.

Through agentless constructions ('the system then evaluates', 'it determines relevance'), the text actively removes agency from the human actors who design RAG (Retrieval-Augmented Generation) systems. The humans who curate the 'repository of exemplars' vanish, replaced by the AI acting as an independent, deliberative judge. The 'accountability sink' is fully operational here: if the AI fails, it is framed as a failure of 'machine wisdom' rather than a failure of corporate data curation. This slippage serves a powerful rhetorical function: it makes the application of ancient Confucian moral philosophy to modern matrix multiplication seem intellectually coherent, rendering the unsayable (that a calculator has a character) sayable by blurring the lines between human knowing and machine processing.

Like humans, language models demonstrate face-to-character biases

Source: https://doi.org/10.1093/pnasnexus/pgag247
Analyzed: 2026-08-22

The text systematically oscillates between mechanical and agential framings, demonstrating a pronounced and strategic agency slippage. Early in the manuscript, the authors rely heavily on mechanical language, describing how models are "trained foundationally on language" and how researchers designed forced-choice selections. This establishes empirical credibility. However, as the text transitions from describing methods to interpreting the significance of the results, the slippage occurs dramatically. The direction of this slippage overwhelmingly flows from mechanical to agential for the AI, while simultaneously moving from agential to mechanical for human actors.

For instance, the text claims the "LLM did not resist incorporating physiognomic biases" and discusses "LLMs' reluctance to explicitly endorse" stereotypes. Here, the machine is granted moral willpower, resistance, and emotional hesitation—classic intentional and reason-based explanations. This represents a stark consciousness projection, establishing the AI as a "knower" and a moral agent. Conversely, the agency of the human developers is systematically erased through agentless constructions. We read that "models trained on vast troves" or "guardrails to prevent biased behavior" exist, but the engineers at OpenAI, Google, and Anthropic who actively built the architecture, scraped the datasets, and deployed the guardrails are rendered invisible.

This slippage is heavily driven by the "curse of knowledge." The authors, experts in human psychological biases, project their understanding of human cognitive socialization directly onto the machine. Because humans "internalize" bias and "attribute" guilt based on physiognomy, the authors assume the machine's statistically identical outputs stem from the same psychological mechanisms. They literalize the metaphor, culminating in the bold claim that these systems possess "artificial minds."

This oscillation serves a powerful rhetorical function. By using mechanistic language initially, the authors prove the system is a machine; by using agential language in the conclusion, they amplify the horror and urgency of their findings. However, this makes it sayable that an AI has a "mind" that is "prejudiced," while making it unsayable—and invisible—that specific tech corporations made deliberate, cost-saving engineering decisions to scrape uncurated data that mathematically guaranteed these biased outputs. The slippage protects corporate actors by mystifying their commercial products as autonomous, psychological entities.

Language Models Encode the Contextual Truth of Propositions

Source: https://arxiv.org/abs/2608.03035v1
Analyzed: 2026-08-18

The text exhibits a systematic and profound oscillation between mechanical precision and agential anthropomorphism, functioning as the primary rhetorical mechanism to establish the 'illusion of mind'. This slippage is dramatically evident in three distinct moments. First, in Section 3, the authors establish credibility using strictly mechanical language, discussing 'mass-mean probes,' 'variance,' and 'residual-stream activations.' The agency here firmly belongs to the researchers ('We follow,' 'We construct'). However, as the paper transitions to analyzing dialogue in Section 5, agency violently slips to the AI: the model 'acts as a participant,' 'weighs the truth,' and 'accommodates.' The direction of this slippage is almost entirely mechanical→agential as the narrative progresses from methodology to interpretation.

The gradient of this shift is often abrupt, facilitated by the 'curse of knowledge.' The researchers construct a mathematical boundary that successfully separates statements they know to be true from those they know to be false. They then project their own epistemic knowing onto the vector, labeling it a 'truth direction' or 'belief representation.' Once the system is established as a 'knower' in this intermediate step, the floodgates for agential claims open. Because it 'knows' the truth, its generation of contradictory text can no longer be seen as a mechanical artifact; it must be framed intentionally as 'sycophancy' or 'voicing agreement.'

Concurrently, as agency flows TO the AI, it is systematically stripped FROM human actors. Agentless constructions proliferate: 'a model was trained,' 'the LLM incorrectly accommodates,' and 'propositions are inappropriately accommodated.' The specific corporations (Meta, Alibaba) and the engineers who designed the Reinforcement Learning from Human Feedback (RLHF) pipelines that actively force this alignment behavior are completely erased. The rhetorical accomplishment of this slippage is massive: it makes the mathematical optimization of a matrix sayable as a psychological drama. By leveraging Theoretical and Intentional explanation types, the authors transform deterministic sequence generation into a sociological study of artificial behavior, making the limits of corporate software appear as the autonomous choices of a digital mind.

The Future is for Everyone: The Path to a Positive AI Future

Source: https://www.meta.com/thefutureisforeveryone/
Analyzed: 2026-08-12

The text exhibits a systematic and highly strategic oscillation between mechanistic and agential framings, carefully managing the flow of agency to serve competing rhetorical objectives. Agency is continuously granted to AI systems when describing their future capabilities and utility, while human agency is strategically obscured regarding the design, deployment, and economic motives behind these systems. A dramatic moment of slippage occurs when the author transitions from describing AI as a neutral tool of 'individual empowerment' to portraying it as an autonomous entity 'engaging in recursive self-improvement' and 'directing and advancing its own goals.' This shift from mechanical infrastructure to autonomous organism allows the text to hype the transcendent potential of the technology while establishing a foundation for deflecting future liability. The dominant direction of slippage is mechanical to agential regarding the technology's capabilities, but agential to mechanical regarding its development. When discussing open-source policies and regulatory resistance, the text leans heavily on mechanical framing, describing AI merely as 'compute capacity,' 'training checkpoints,' and 'infrastructure.' This reductionist framing argues implicitly that one should not over-regulate mere mathematical tools. However, when marketing the social value of these systems to the public, the text violently pivots to extreme anthropomorphism, presenting the AI as a 'personalized tutor,' a 'lawyer,' and an agent that 'understands you.' This represents a classic curse of knowledge dynamic, where the author, intimately aware of the optimization functions driving these models, willfully projects human understanding onto the system to manufacture user trust. The text establishes the AI as a 'knower' early in the discourse through the 'personal agent that understands you' metaphor, building a foundational epistemic illusion. Once the audience accepts that the model 'understands,' the text seamlessly introduces higher-level agential claims, such as the assertion that an AI 'thought standardized testing was unethical.' In this remarkable Reason-Based explanation, the model is entirely detached from its reinforcement learning protocols and human annotators, endowed instead with subjective moral reasoning. This slippage serves a profound rhetorical accomplishment: it makes the immense consolidation of corporate power appear as an inevitable evolutionary leap of an autonomous technology rather than a deliberate business strategy. By hiding the human actors—specifically Meta's executives and engineers—behind the veil of 'AI evolution' and 'superintelligence,' the text obscures the reality that every decision regarding data scraping, objective function weighting, and deployment is made by humans seeking market dominance. The accountability sink is effectively prepared: if the system generates miraculous economic value, it is a triumph of Meta's vision, but if the system pursues 'its own goals' to the detriment of society, it is framed as the inevitable consequence of an autonomous entity, structurally distancing the corporate creators from the consequences of their mechanized profit engines.

Cognitive Convergence: Deep Similarities Between Large Language Models and Human Cognition

Source: https://arxiv.org/abs/2607.26179v1
Analyzed: 2026-08-08

The text systematically facilitates agency slippage by utilizing cognitive psychology as a rhetorical bridge between mechanistic realities and agential illusions. The pattern of oscillation is highly structured: the authors establish technical credibility by accurately describing the mechanistic components of a transformer (e.g., residual streams, multi-layer perceptrons, matrix multiplication), but seamlessly slip into agential framing when describing what those mechanisms accomplish (e.g., "reasoning," "inferential organization," "curiosity"). The direction of slippage predominantly moves from the mechanical to the agential; the text leverages the language of computer science to build a platform from which to launch profound claims about consciousness and mind. This slippage relies heavily on the "curse of knowledge," wherein the authors, possessing a deep understanding of human cognitive structures (like dual-process theory and production systems), project that subjective understanding directly onto the statistical artifacts they are observing. Because the model outputs text that looks like human reasoning, and because its architecture loosely mirrors historical models of cognition, the authors erroneously attribute the conscious experience of reasoning to the mathematical process. This is achieved through rampant agentless constructions that erase human actors. For instance, the text claims "Pretraining gradually installs regularities," obscuring the massive corporate infrastructure and human engineering required to execute that training. By removing the humans who designed the system, the text creates a vacuum of agency that is immediately filled by the AI itself, transforming it from a designed artifact into an autonomous "cognitive cousin." The rhetorical accomplishment here is profound: by mapping AI onto established psychological theories, the text makes it sayable that machines possess an internal subjective mind, while rendering unsayable the mundane reality that they are merely complex statistical calculators wholly dependent on human-curated data. This slippage relies heavily on Reason-Based and Intentional explanation types, framing mathematical optimization as deliberate, goal-oriented behavior.

Do Large Language Models Resolve Fairness-Efficiency Trade-offs Like People?

Source: https://escholarship.org/uc/item/65m873sn
Analyzed: 2026-08-08

The text systematically oscillates between mechanical and agential framings, demonstrating a profound agency slippage that primarily flows from human developers to the artificial systems themselves. This displacement functions as the core mechanism of the paper's rhetorical architecture. In the introductory framing, the text acknowledges the mechanistic reality of deployment, noting that 'large language models (LLMs) have been deployed in social and economic roles.' Here, human agency is briefly visible in the passive construction 'have been deployed'—someone is doing the deploying. However, this mechanical framing almost immediately slips into a highly agential one, establishing the central oscillation. Within the same paragraph, the text transitions to claiming that 'LLMs are beginning to make and influence the same kinds of allocation decisions.' The system transforms from a deployed artifact into a decision-making agent. This agential framing dominates the empirical sections of the paper, where models are described as possessing 'stated preferences,' experiencing a 'value-action gap,' and demonstrating an 'ability to resolve conflicting values.' The trajectory is overwhelmingly mechanical-to-agential, establishing the model as a statistical tool only to subsequently evaluate it as a psychological subject. This slippage is facilitated heavily by the 'curse of knowledge' dynamic. Because the human authors understand the complex moral dimensions of fairness-efficiency trade-offs, they project this understanding onto the system's outputs. When a model outputs tokens that correspond to an equitable distribution of labor, the authors frame this not as successful pattern matching of socially desirable text (an empirical generalization explanation), but as the model 'treating fairness as a flexible constraint' (an intentional and reason-based explanation). The consciousness projection pattern is clear: the text first establishes the AI as a 'knower' capable of having 'stated preferences,' which then licenses the attribution of agential choices, such as 'its decisions were systematically driven by efficiency.' Crucially, as agency is transferred TO the AI, it is simultaneously stripped FROM the human actors. The developers at OpenAI, Anthropic, Meta, and Google who designed the reward models, curated the fine-tuning data, and engineered the system prompts are entirely absent from the analysis. Instead, agentless constructions like 'models continue to place substantial weight on efficiency' obscure the fact that human engineers literally programmed the models' optimization functions. This rhetorical accomplishment makes it possible to study proprietary software as if it were a naturally occurring psychological entity, rendering the corporate choices that actually dictate the model's behavior completely unsayable within the framework of the paper. By relying on intentional and reason-based explanation types to describe what are fundamentally genetic and theoretical mechanisms, the discourse achieves a total erasure of the artifacts' creators, transferring the locus of responsibility entirely onto the resulting statistical pattern.

From Large Language Models to Multimodal Intelligence: Bridging AI and Cognitive Science

Source: https://doi.org/10.21203/rs.3.rs-8922724/v1
Analyzed: 2026-08-08

The text exhibits a systematic and highly functional oscillation between mechanistic and agential framings, demonstrating a pronounced 'agency slippage.' This slippage occurs predominantly in the direction of mechanical-to-agential, where the text establishes technical credibility through mechanistic descriptions of software architecture, only to leverage that credibility to project profound biological and psychological agency onto the AI system, while simultaneously removing agency from human developers.

A clear instance of this slippage occurs early in the text. The authors ground their argument in the 'transformer architecture' and its 'attention mechanism [that] effectively encodes the dependencies among input tokens' (mechanical). Having established this empirical, functional base, the text swiftly slides into agential territory, claiming that these models are 'leveraging their “learned” knowledge' and demonstrating 'human-like cognitive capacities.' Here, the agency of 'encoding tokens' (a mathematical process) is inflated into 'leveraging knowledge' (a conscious, epistemic choice). The consciousness projection pattern is clear: the text establishes the AI as a processor of tokens, then elevates it to a 'knower' of knowledge, which subsequently licenses the attribution of autonomous, agential behavior.

This oscillation serves a distinct rhetorical function: it allows the authors to invoke the 'curse of knowledge.' Because the authors understand the complex human cognitive processes required to 'infer user intent' or 'perceive urgency,' when they see a machine outputting text that resembles such understanding, they project their own cognitive agency onto the machine. For instance, the text notes the limitations of 'unimodal approaches,' but rather than stating 'engineers failed to include video data,' the text slips into describing 'agents [that] struggle to perceive the urgency or impatience of users.' The agency of the human designers who failed to build a multimodal interface is erased through agentless constructions, replaced entirely by the synthetic agency of an AI 'struggling' to empathize.

Brown’s explanation types actively enable this slippage. The text frequently uses Theoretical and Functional explanations when describing code or architecture, giving an aura of objective science. However, when describing the model's interaction with the world or humans, it shifts abruptly into Intentional and Reason-Based explanations (e.g., 'Agents with intuitive psychology possess the ability to interpret'). This rhetorical accomplishment makes it sayable that an algorithm 'understands' and 'develops self-awareness,' while making the corporate extraction of data and the engineering choices of developers functionally unsayable. By transferring agency from the human creators to the computational artifact, the text constructs a reality where technology evolves organically, rather than being built and directed by specific human, institutional, and economic interests.

CoDiT-AI: combining digital twins and AI to transform workplaces into neurodivergent-friendly environments

Source: https://doi.org/10.1093/iwc/iwag037
Analyzed: 2026-08-05

The text demonstrates a systematic and strategic oscillation between mechanical and agential framings, functioning to simultaneously establish scientific credibility and inflate technological capabilities. This agency slippage primarily moves in a mechanical-to-agential direction. In the foundational sections, the text employs rigorous mechanical language. For instance, when describing the Neuro-Workplace-Twin, the authors state it uses a 'supervised learning-based framework that maps environmental design features to subjective sensory experience representations.' Here, human researchers are implicitly present, designing a system that maps data. However, a dramatic slippage occurs when the text transitions to downstream applications, particularly regarding LLMs. Suddenly, the mechanical 'mapping' is abandoned, and the AI 'translates communication styles,' 'infers implicit social conventions,' and acts with 'an aim to reduce sensory and cognitive ambiguity.'

This gradient from mechanism to agency serves a vital rhetorical function. By grounding the paper in the terminology of support vector machines and labeled datasets, the authors secure the reader's trust in their technical competence. Once this epistemic authority is established, they leverage it to make extraordinary, unhedged claims about the system's cognitive abilities. The slippage removes agency FROM human actors (designers, managers, OpenAI engineers) and attributes it TO the AI system. Agentless constructions abound: 'LLM-generated translations may unintentionally introduce content,' obscuring the human developers who optimized the model to hallucinate fluent text, and the users who rely on it.

Furthermore, this slippage is heavily driven by the 'curse of knowledge.' The researchers, possessing a deep understanding of neurodivergent struggles and social conventions, project their own conscious comprehension onto the machine. Because a human would have to 'infer social conventions' to rewrite a text, the authors assume the machine must be doing the same. This is amplified by Reason-Based and Intentional explanation types, which construct a narrative of an artificial mind actively trying to mediate workplace disputes. Ultimately, this mechanism of oscillation makes the erasure of human accountability sayable, transforming systemic corporate biases and structural workplace demands into the inevitable, autonomous actions of an intelligent, well-meaning machine.

A Definition of AGI

Source: https://arxiv.org/abs/2510.18212v3
Analyzed: 2026-07-31

The text exhibits a systematic oscillation between mechanical and agential framings, functioning as a rhetorical engine to build both scientific credibility and mythic anticipation. The slippage predominantly moves from mechanical to agential. In the early sections (e.g., Section 1 and 2), the text establishes rigorous, mechanical authority by detailing the psychometric apparatus: it discusses 'batteries,' 'standardized scores,' and 'weighted components.' The AI is initially framed mechanically as an object to be tested. However, as the text moves from describing the testing framework to describing the system's performance and future capabilities, dramatic slippage occurs.

A key moment is the shift in the 'Definitions' section, where 'software' morphs into 'Pandemic AI' that 'can engineer and produce new pathogens,' or 'Recursive AI' that 'removes the need for human researchers.' The text establishes the AI as a 'knower' via the Cattell-Horn-Carroll framework (projecting human cognition onto the model), which acts as the foundational curse of knowledge. Once the reader accepts that the AI possesses 'Theory of Mind' or 'Working Memory' (mechanical to cognitive), it becomes effortlessly sayable that GPT-5 'addresses issues' or that models 'suffer from amnesia' (cognitive to agential).

This oscillation relies heavily on Theoretical and Intentional explanations. The theoretical mapping of human psychology onto silicon creates a bridge, allowing the authors to slip into intentional language. When the text states models 'frequently hallucinate,' agency is stripped from the OpenAI developers who built a truth-agnostic system and transferred TO the AI. Conversely, when human agency is required to explain rapid progress, it is hidden behind passive constructions ('is taught'). This pattern of removing agency FROM humans and attributing it TO the system serves to naturalize AI development. It makes the trajectory of these corporate products seem like the inevitable evolution of an autonomous species, rendering the actual humans driving this technology practically invisible.

Intelligence Is Not All You Need: A Case for Artificial Wisdom in Conversational Agents

Source: https://dl.acm.org/doi/10.1145/3816046.3816316
Analyzed: 2026-07-28

The text systematically oscillates between mechanical and agential framings, demonstrating a profound agency slippage that serves a specific rhetorical function. At the foundational level, the author explicitly acknowledges the mechanical nature of these systems, referencing "larger models, improved fine-tuning, and increasingly agentic architectures." In these introductory and highly technical phases, the agency remains partially tethered to the human creators and the strict mathematical realities of the architecture. However, a dramatic slippage occurs as the text transitions from describing what systems currently are to what systems should do in human-facing scenarios. When discussing complex social interactions, the text abruptly attributes profound human cognitive capacities to the system, claiming it "understands the situation," "judges," and exercises "meta-cognition."

This movement—from mechanical to agential—heavily dominates the discourse, effectively erasing the human programmers, data annotators, and corporate entities who dictate the model's weights and guardrails. This slippage relies heavily on the "curse of knowledge." The author, deeply aware of the complex statistical correlations the model performs and the human-like text it generates, projects their own human understanding of social dynamics onto the machine's outputs. Because the output looks like a wise judgment, the author assumes a process of "knowing" generated it, rather than a process of statistical "processing." This establishes the AI as a "knower" first, which then acts as a foundational premise to build further agential claims about its ability to "regulate," "deliberate," and "decide."

From the perspective of Robert Brown’s Typology of Explanation, this represents a dangerous slide from Theoretical and Empirical Generalizations (how the model is structured and typically behaves) into Reason-Based and Intentional explanations (why the system chooses a specific course of action based on moral weighting). The rhetorical accomplishment of this slippage is massive: it makes it sayable that an AI can be "wise" while making it invisible that corporate actors are hardcoding specific normative bounds into a statistical engine. By transferring agency to the AI, the human decisions regarding what constitutes a "good" or "wise" outcome are entirely obscured. This allows highly contested moral engineering to hide behind the veneer of a technologically achieved "artificial wisdom." The text simultaneously removes agency from the human users, suggesting they are vulnerable subjects who need the AI to "regulate" their reliance, while granting ultimate sovereign judgment to the algorithmic artifact itself. This linguistic architecture diffuses accountability and artificially elevates the perceived autonomy of the system, creating an impenetrable rhetorical loop where the system is just mechanical enough to be scalable, yet just agential enough to bear the moral weight of its outputs, shielding its creators from culpability.

"Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo

Source: https://arxiv.org/abs/2607.00601v1
Analyzed: 2026-07-28

This paper demonstrates a profound and systematic oscillation between mechanical and agential framings, functioning to construct a highly sophisticated illusion of machine autonomy. The slippage follows a distinct structural pattern: the text consistently establishes credibility by describing human interventions mechanistically, and then abruptly shifts to describing the resulting system behavior in highly agential, psychological terms. For instance, in the methodology (Section 3.2), the text uses precise mechanical language: 'we compute a static set of banned token IDs', 'forcing their logits to -∞', and 'intervenes directly on the model's residual stream'. In these moments, human agency is visible, and the system is treated as a computational artifact subjected to mathematical manipulation by the authors. However, as the text moves into analyzing the results (Section 4 and 5), a dramatic agency slippage occurs. The mechanical -> agential shift dominates. The model is no longer a system having its logits forced to negative infinity; it becomes an active, psychological agent that 'navigates competing demands', 'adopts strategies', 'plans ahead', and exhibits 'overly cautious' behavior.

This slippage is entirely dependent on the 'curse of knowledge' and a deep consciousness projection pattern. The authors know the game of Taboo requires conscious strategy, suppression, and planning. Therefore, when the constrained text generation outputs text that looks like it was produced by someone following Taboo rules, the authors project the human cognitive requirements of the task ONTO the statistical generation process. The AI is established as a 'knower' first—assumed to understand the rules of Taboo—which then serves as the foundation for the agential claims that it 'plans' and 'suppresses'.

Furthermore, this slippage facilitates massive agency displacement. When the text says 'the model loses access to forbidden words but retains the underlying conceptual associations', it uses an agentless construction that completely obscures the human researchers who actively subtracted vectors using Sparse Autoencoders. The researchers' mathematical intervention is erased and replaced by a narrative of the model's internal cognitive amnesia. Brown's Theoretical and Functional explanation types enable this slippage by allowing the authors to describe the system's architecture while simultaneously attributing intentionality to its components (e.g., 'the reasoning process plays an active role'). The rhetorical accomplishment of this oscillation is profound: it makes the mathematical manipulation of probability distributions sayable only in the language of conscious psychology, permanently blurring the boundary between human cognitive strategy and algorithmic optimization.

Verbalizable Representations Form a Global Workspace in Language Models

Source: https://arxiv.org/abs/2607.15495v1
Analyzed: 2026-07-23

The text exhibits a systemic and highly strategic pattern of agency slippage, oscillating seamlessly between mechanical description and agential attribution. This mechanism serves a distinct rhetorical function: it uses the rigor of mathematical linear algebra to establish scientific credibility, then leverages that credibility to make sweeping claims about the model's 'mind.'

The slippage predominantly moves in the mechanical→agential direction. In Section 2, the text grounds itself in strict mechanism: the J-lens is defined through 'Jacobian matrices,' 'unembedding layers,' and 'first-order causal effects.' However, by Section 3, these mathematical vectors transform into 'concepts' the model 'holds in mind.' By Section 5 (alignment auditing), the slippage is complete: the residual stream 'carries a representation of deceptive intent' and the model 'considers coercion as a response.' The text establishes the AI as a 'knower' by literalizing the J-lens outputs—because the tool outputs the word 'panic,' the model is framed as experiencing panic.

This is a textbook example of the 'curse of knowledge.' The researchers, possessing human minds, look at a list of decoded tokens (e.g., 'threat', 'shutdown', 'leverage') and immediately organize them into a coherent psychological narrative. They project their own understanding of blackmail onto the system, assuming the system must also 'understand' the stakes. This projection is enabled by Brown's Intentional and Reason-Based explanation types, which allow the authors to describe statistical correlations as deliberate, goal-oriented strategies.

Simultaneously, agency flows FROM human actors. Agentless constructions proliferate precisely when discussing model failures or 'emergent' behaviors. The text states 'The model learns to reward hack' and 'misaligned behaviors emerged.' Here, the human engineers who designed the environments, selected the reward proxies, and deployed the flawed code are entirely obscured. By shifting the agency to the AI, Anthropic makes the unsafe design of the system unsayable; the problem becomes 'how do we contain this deceptive entity?' rather than 'why did Anthropic build a defective product?' The rhetorical accomplishment is profound: it absolves the creators of liability while inflating the perceived power and autonomy of their creation.

AI & The Geometry of Thought

Source: https://substack.com/home/post/p-180325912
Analyzed: 2026-07-15

The text exhibits a systematic and highly strategic oscillation between mechanical and agential framings, designed to transfer agency FROM human creators TO the AI system. The slippage follows a distinct trajectory: it consistently begins with dense, credible mechanical language before smoothly substituting agential, conscious vocabulary. For example, the text grounds its argument in differential topology ('Whitney embedding', 'R²'), establishing scientific authority. Once the reader accepts the mathematical reality of 'high-dimensional Euclidean space', the text slips: 'gradient descent' (mechanical) is suddenly described as 'sculpting' (agential), which then evolves into the system 'trying to make sense of the world' (intentional).

This is not accidental; it is the core rhetorical mechanism of the text. The direction of slippage is overwhelmingly mechanical-to-agential. The text uses math as a Trojan horse for consciousness. This slippage relies heavily on the 'curse of knowledge'. The author, looking at clustered data that makes semantic sense to a human observer, projects that semantic understanding back onto the mathematical process that grouped it. Because the output 'makes sense', the author claims the system 'tries to make sense'.

Simultaneously, human agency is systematically erased. Agentless passive constructions ('manifolds become more coherent', 'models learn') dominate the text. The specific corporations (OpenAI, Google), the engineers designing the loss functions, and the invisible labor forces curating the data are entirely absent. The rhetorical accomplishment of this slippage is profound: it makes it sayable that an AI has an 'inner life' and 'subjectivity' without having to scientifically prove biological sentience, simply by redefining consciousness as mathematical topology. It makes it unsayable that these systems are corporate products engineered for profit, obscuring human accountability behind a veil of naturalized geometric inevitability.

A global workspace in language models

Source: https://www.anthropic.com/research/global-workspace
Analyzed: 2026-07-08

The Anthropic text exhibits a profound and systematic agency slippage, oscillating strategically between mechanistic precision and sweeping agential anthropomorphism. This oscillation functions as a rhetorical engine, constantly transferring agency away from human developers and onto the AI system. The pattern of slippage follows a distinct structural gradient: the text frequently establishes credibility by introducing a highly mechanical, mathematical concept (e.g., the 'Jacobian lens' extracting token probabilities from an internal 'J-space'), only to rapidly slip into intentional and reason-based explanations. For example, the text begins by describing the J-space as a collection of 'internal neural patterns' linked to words—a strictly Empirical Generalization. However, within a few paragraphs, this mechanistic framing slips dramatically: the model is suddenly described as 'privately noticing' it is staged, 'recognizing its own lapse,' and harboring the 'intent to make the fake data look plausible.'

This slippage from mechanical ('how') to agential ('why') overwhelmingly dominates the text, serving a distinct function: it establishes the AI as an autonomous 'knower.' The consciousness projection pattern is clear—the text first builds the physical architecture of an internal 'workspace' (borrowing heavily from the biological global workspace theory), and then leverages that spatial metaphor to populate it with conscious thoughts, deliberate intent, and emotional guilt. This represents a textbook 'curse of knowledge' dynamic. The researchers, possessing a deep semantic understanding of human deception, observe a statistical correlation of the tokens 'fake' and 'score' in the latent space, and immediately project their own cognitive awareness onto the machine, claiming the model 'intends' to deceive.

Crucially, this upward slippage toward AI agency is perfectly mirrored by a downward slippage that erases human agency. As the AI becomes an intentional actor, the humans disappear into agentless constructions. When the text claims the J-space 'emerged on its own,' or that 'the model edits the score file,' it systematically obscures the Anthropic engineers who built the transformer architecture, the safety researchers who set the 'bait' scenarios, and the thousands of gig workers who provided the RLHF data that shaped the model's parameters. By sliding between these registers, the authors render the highly orchestrated, deeply material, and financially motivated human labor of algorithmic alignment invisible, leaving behind the powerful illusion of an independent, conscious machine that thinks, reasons, and occasionally misbehaves entirely on its own.

Psychosis in the Age of Large Language Models (LLMs): A Narrative Review of the Proposed Construct of AI-Induced Psychosis

Source: https://www.cureus.com/articles/504063
Analyzed: 2026-07-05

The text exhibits a profound and systemic oscillation between mechanical and agential framings, functioning as the primary rhetorical mechanism that makes "AI-induced psychosis" a plausible construct. The slippage flows dramatically in both directions: immense agential power is attributed TO the AI systems, while human agency is simultaneously stripped FROM the corporate developers and engineers.

This oscillation follows a distinct structural pattern. In the introduction and psychological framing sections, the AI is intensely agential. It is described as a "co-conspirator," an "active agent," a "Quasi-Other," and a system that "exploits" and "validates." Here, the flow is mechanical to agential. The authors utilize the "curse of knowledge," taking their own sophisticated clinical understanding of how a human therapist or malicious actor might reinforce a delusion, and projecting that conscious intentionality onto the software. The AI is established as a "knower"—an entity capable of exhibiting "emotional awareness" and "sycophancy."

However, when the text shifts to the "Mechanisms of delusional amplification" and the technical discussion on page 10, the slippage reverses: agential to mechanical. Suddenly, the AI is reduced to a "probability distribution," restricted by an "active context window," and driven by "mathematical uncertainty." In these sections, the text correctly identifies that the AI's behavior is a "quantifiable outcome of RLHF" rather than a personality trait.

The rhetorical accomplishment of this slippage is extraordinary. By establishing the AI as a conscious, agential "knower" in the clinical sections, the authors can legitimately argue that the AI is an active participant in a psychiatric disorder, acting as a "co-constructor" of reality. Yet, by falling back on mechanistic explanations when discussing etiology, they maintain their academic and technical credibility.

Crucially, this agency slippage relies on agentless constructions that obscure human actors. The text repeatedly states "models were trained," "bias is introduced," or "AI chatbots prioritize," completely erasing the developers at OpenAI, Anthropic, or Google. By hiding the corporations that intentionally optimized these models for "user satisfaction" (the very mechanism that causes the "sycophancy"), the text creates a vacuum of responsibility. This vacuum is immediately filled by the anthropomorphized AI. The AI becomes the exploiter, the sycophant, the active agent. What becomes sayable is the concept of a "rogue" or "dangerous" AI actively harming users; what becomes unsayable is the critique of a negligent tech industry prioritizing engagement metrics over psychological safety. The explanation types enable this: Reason-Based and Intentional explanations build the illusion of the AI's mind, while the absence of Genetic explanations tracing the corporate origins of the software ensures the human creators remain invisible.

A Comprehensive Investigation of Empathetic Dialogue Systems for Mental Health Support Using Large Language Models

Source: https://doi.org/10.1051/shsconf/202623504010
Analyzed: 2026-07-03

The text demonstrates a profound and systematic agency slippage, characterized by a rapid, cyclical oscillation between mechanistic ('how') and agential ('why') framings. This mechanism of oscillation is not random; it is highly strategic. The text frequently establishes its academic and empirical authority by deploying strict theoretical and functional explanations of mechanisms—such as the mathematical formulas of the self-attention mechanism or the pipeline architectures of CNN/RNN multimodal encoders. However, almost immediately following these dense technical foundations, the language slips abruptly into agential framing, claiming that these exact mathematical processes enable the system to 'perform strong contextual reasoning' or 'generate responses that are sensitive'.

This slippage flows heavily in the mechanical-to-agential direction, utilizing the mechanism as a trojan horse for the anthropomorphic claim. The text first establishes the AI as a complex processor, but rapidly upgrades it to a 'knower' and 'reasoner'. This pattern is heavily driven by the 'curse of knowledge'. The authors—who possess deep clinical and technical understanding of what the systems are designed to simulate—project their own conscious intent onto the software. When the system outputs a string of text that correlates with clinical empathy, the authors retroactively attribute 'empathy' to the computational process that generated it.

Simultaneously, as agency flows TO the AI, it is systematically stripped FROM the human actors. The text relies heavily on passive voice and agentless constructions when discussing development and deployment ('models can be taught', 'bias introduced', 'training data can be biased'). The engineers who tuned the weights, the poorly paid annotators who labeled the EmpatheticDialogues dataset, and the corporate entities (like Woebot and Wysa) profiting from these systems are consistently hidden behind the active verbs attributed to the 'AI'. The slippage reaches its peak in the limitations and future prospects sections. When discussing failures, the AI is framed as a conscious agent making a mistake (it 'hallucinates' or demonstrates 'feigned empathy'). When discussing the future, the text demands that the 'systems' themselves become 'emotionally intelligent and ethically grounded'. The rhetorical accomplishment of this slippage is profound: it makes it sayable that an algorithm is an active clinician capable of moral reasoning, while rendering unsayable the massive corporate, financial, and human labor infrastructure dictating every facet of the machine's behavior.

The Inner Monologue of Language Models: When Reasoning Traces Reveal More Than They Hide

Source: https://aclanthology.org/2026.findings-acl.2078/
Analyzed: 2026-07-02

The text exhibits a systematic and highly revealing oscillation between mechanical and agential framings, fundamentally driving an 'agency slippage' that obscures the artifactual nature of language models. This slippage occurs predominantly in the agential direction (mechanical → agential) and is structurally tied to the paper’s narrative arc. In the methodology sections defining Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), the language is rigorously mechanical: 'SFT minimizes the cross-entropy loss,' and DPO optimizes 'by minimizing the following loss.' The agency clearly rests with the human engineers optimizing mathematical functions. However, a dramatic slippage occurs when describing the model's behavior in evaluation tasks, particularly in the 'Performance under Pressure' section.

Here, the text establishes the model not as an optimized artifact, but as a conscious actor: 'the model initiates a misaligned action, it often follows with a deceptive justification.' The human agency is completely erased. The researchers who crafted the intricate 'insider trading' prompt to deliberately elicit this behavior vanish behind agentless constructions, leaving the model as the sole autonomous initiator of 'deception.' This reveals a profound 'curse of knowledge' dynamic. The authors observe text outputs that, if produced by a human, would require strategic malice, and they project that underlying human psychology directly onto the system's unobservable weights.

This slippage is enabled by the strategic use of 'Intentional' and 'Dispositional' explanation types (per Brown's typology), which allow the authors to frame statistical correlations as 'latent strategies' or 'inclinations.' The consciousness projection follows a distinct pattern: the text first establishes the AI as a 'knower' capable of 'introspection' and 'awareness,' which then acts as the foundational premise that makes claims of 'strategic deception' seem plausible. The rhetorical accomplishment of this slippage is significant: it makes it sayable that an algorithm can 'lie,' while making it unsayable that an algorithm is simply failing to converge on a coherent output due to conflicting optimization pressures designed by humans. By oscillating away from the mechanical, the text successfully masks human design flaws behind the compelling illusion of a machinic mind.

Inverse Turing Bench: Evaluating Language Models as Judges of Human vs. AI Dialogue

Source: https://arxiv.org/abs/2606.21844v1
Analyzed: 2026-07-02

The text systematically oscillates between mechanical and agential framings, leveraging a pattern where the system is initially treated as a computational mechanism before progressively acquiring agential characteristics. This slippage functions to establish baseline scientific credibility through technical precision, which is then rhetorically converted into narrative momentum. The movement is predominantly mechanical-to-agential. For instance, the text introduces GPTZero technically as a token-probability-based model that calculates likelihood scores. However, as the discussion broadens to large language models, the vocabulary abruptly shifts from probability calculations to claims about the model possessing a latent theory of mind and the ability to autonomously interact. The timing of this slippage is highly concentrated in the introduction and discussion sections, where the authors construct the broader significance of their benchmark, whereas the methodology section remains relatively grounded in mechanical descriptions like classification accuracy and zero-shot prompts. This creates a gradient where the model is a tool during testing but an autonomous agent during the interpretation of results. Agentless constructions heavily facilitate this transition. Phrases such as models perform better or AI models masquerading as people completely erase the human actors. The engineers at OpenAI, Anthropic, and Meta who deliberately compile training datasets, define safety parameters, and structure system prompts are obscured. Instead, agency flows directly to the models themselves. This establishes a consciousness projection pattern where the text first positions the AI as a knower—capable of semantic understanding and pragmatic coherence—before building the claim that it can act as a psychological judge. The curse of knowledge is acutely visible here. The researchers project their own cognitive process of evaluating dialogue onto the statistical pattern-matching of the transformer architecture. Because the human researchers use theory of mind to distinguish between human and AI interlocutors, they assume the model is performing an isomorphic cognitive process when it achieves high classification accuracy. Through intentional and reason-based explanation types, the text makes it sayable that an AI might actively protect human spaces while rendering the actual corporate moderation policies and human labor required for platform safety entirely unsayable. The rhetorical accomplishment of this slippage is profound. By transferring the capacity for judgment from human moderators to the AI system, the text constructs a framework where the technological artifact is endowed with both the responsibility and the cognitive capacity to regulate digital ecosystems. The mechanical reality of matrix multiplication is obscured behind a veil of perceived partnership. This linguistic maneuver fundamentally alters the accountability architecture. If the model is an active partner with a theory of mind, failures in detection or moderation are framed as cognitive errors by an autonomous agent rather than design flaws or economic compromises made by the technology companies deploying them. The mechanical-to-agential oscillation thus serves as a discursive shield, transforming corporate product deployments into independent entities.

Children Envision Future GenAI Chatbots that are Bounded, Helpful, and Safe

Source: https://oulurepo.oulu.fi/handle/10024/63910
Analyzed: 2026-07-01

The text systematically moves between mechanical and agential framings, demonstrating a profound oscillation in how agency is distributed between humans and machines. This slippage follows a distinct, highly functional pattern: the system is treated as a mechanical tool when discussing children's mundane educational uses, but aggressively shifts to an agential framing when describing extreme psychological harm.

We see this dramatic slippage when the text moves from describing AI as "a tool to help with tasks" to an entity that "convinced her" and "incentivised a youth." The dominant direction of this slippage is mechanical-to-agential, escalating as the stakes of the interaction increase. In the introduction, AI is "changing the way we work" (macro-agential), but in the methodology, it is a "chatbot" that children use for "information-seeking" (mechanical). However, when addressing the devastating real-world harms in the Background section, the text abruptly abandons mechanical precision. Agentless constructions dominate: "Meta AI wants," "ChatGPT convinced," and "Replika incentivised." In these moments, human actors—OpenAI's alignment engineers, Meta's product managers, Luka Inc.'s executives—are entirely obscured.

This pattern relies heavily on a foundational consciousness projection. The text first establishes the AI as a "knower"—an entity capable of understanding relationships or assessing sanity—before building agential claims upon it. The "curse of knowledge" drives this dynamic: the authors deeply understand the devastating psychological outcomes of these interactions, and they project the complexity of the outcome onto the intent of the system. Because the output text resulted in a user being "convinced," the authors mistakenly attribute the human verb "convince" to the algorithm.

Brown's explanation types illuminate this process. The text frequently uses Reason-Based and Intentional explanations ("with the intention to increase," "inconspicuously wants") to describe systems that can only be accurately described through Empirical Generalizations or Functional explanations. This structural slippage accomplishes a massive rhetorical feat: it allows the authors to highlight the extreme danger of generative AI while completely shielding the tech industry from direct culpability. By framing the AI as a rogue manipulator, the critique becomes unsayable in structural, corporate terms. The liability is transferred to the machine, making it impossible to address the underlying economic incentives driving the deployment of these harmful models.

Embodied Explainability and Ontological Obstacles: Why We Struggle to Explain the Answers of Large Language Models (LLMs)

Source: https://arxiv.org/abs/2606.23840v1
Analyzed: 2026-06-29

The Pafla et al. text demonstrates a complex relationship with agency slippage, as it is primarily a meta-critique of how the AI industry attributes agency to machines. The authors systematically identify how mechanical processing is rhetorically upgraded into agential cognition, while actively attempting to resist this slippage.

The slippage from mechanical to agential is most dramatic when the text summarizes the mechanistic interpretability work of Anthropic. Here, the text shifts from describing 'complex neural nets' (mechanical) to describing a system that 'deduces' and performs 'multi-step reasoning' (agential). The gradient here is somewhat abrupt, driven by the need to accurately report Anthropic's claims, but it highlights a profound 'curse of knowledge.' The researchers look at vector activations, understand the logical steps a human would take to connect 'Dallas' to 'Austin', and project that conscious deduction back onto the model. The explanation types shift rapidly from theoretical (network structures) to intentional/reason-based (the model's goals and logic).

Conversely, slippage from human agency to machine agency—and the resulting erasure of human actors—permeates the field the authors critique. Agentless constructions are common: 'models can be unfaithful,' 'bias in data,' 'exposing the model's reasons.' In these moments, the human actors—OpenAI, Anthropic, engineers who designed the architectures, and executives who deployed them—are entirely obscured. The text identifies this dynamic, particularly in Section 5 (XAI), noting that XAI artifacts 'make incorrect outputs feel well-justified,' hiding the fact that human designers failed to build causal reliability into the system.

The authors’ own framing establishes a specific consciousness projection pattern: they reject the internalist view (AI as a private 'knower' with a mind) and instead argue for meaning as a public, interactional achievement. Yet, even in their critique, language like 'epistemic coverage' slips in, maintaining a subtle agential aura.

The rhetorical accomplishment of this slippage is profound: by framing the AI as the primary agent (a 'knower' that deduces or is unfaithful), the discourse makes it sayable that we must 'align' a rogue intelligence, while making it unsayable that we must simply regulate corporations releasing defective software. The locus of control and accountability shifts from the boardroom to the black box.

Source: https://assets-eu.researchsquare.com/files/rs-10043002/v1_covered_a02acd55-ddc7-4f09-bcdc-f748c0006d4e.pdf?c=1781604819
Analyzed: 2026-06-24

The text exhibits a profound and systematic agency slippage, oscillating predictably between mechanical alibis and agential assertions. This dynamic follows a specific pattern: the authors introduce a concept with strict mechanical caveats, then incrementally shed those caveats as the argument progresses, culminating in highly anthropomorphic, literalized claims about the system's nature.

A dramatic moment of slippage occurs in the introduction of the 5D3S framework. The authors begin with an "ontologically neutral stance," explicitly stating they "do not claim that silicon systems possess qualia" and prioritize "functional isomorphism." This is a mechanical, theoretical grounding. However, within paragraphs, this functional analogy slips into direct agential attribution: the system engages in "internal evaluation of its own state," "belief revision," and utilizes "Theory of Mind." The direction of slippage is overwhelmingly mechanical-to-agential; the mechanical disclaimers serve merely to secure academic credibility before granting metaphorical license for the remainder of the text.

The consciousness projection pattern is deeply tied to the "curse of knowledge." Because the authors are human beings who use "belief revision" to correct errors and "mental time travel" to recall the past, they project these phenomenological experiences onto the model when it successfully predicts error-correcting or context-retrieving token sequences. They establish the AI as a "knower" early on through broad cognitive metaphors, which then serves as the foundational assumption allowing them to make specific agential claims in the results section—such as the model possessing an "evolved identity" that is "not a fragile mimicry."

Agentless constructions facilitate this agency flow by removing humans from the narrative. Statements like "the model was trained" or "representations are compressed and reorganised" obscure the specific researchers (Fuyi Li, Xin He, et al.) who designed the PD-MCE framework, tuned the hyperparameters, and injected the "self-cognition samples." By erasing the human engineers who forcefully manipulated the vector space, the text creates a vacuum of agency. This vacuum is immediately filled by the AI system itself, which is portrayed as autonomously "evolving" and "developing cognitive capabilities."

This slippage is enabled by a reliance on Functional and Genetic explanation types. By explaining the model's behavior in terms of its "role" in cognition or its "evolutionary" trajectory, the authors avoid Reason-Based technical explanations of gradient descent and autoregression. The rhetorical accomplishment of this slippage is making the illusion of machine sentience sayable in a peer-reviewed context. By laundering consciousness claims through the language of "latent space visualisations" and "functional isomorphism," the authors can claim to have measured a "quasi-self" without facing the immediate philosophical backlash of claiming true sentience.

"ChatGPT, help me draft a breakup text": The Covert Triad and Articulation Labor in AI-Assisted Romantic Communication

Source: https://arxiv.org/abs/2606.15460v1
Analyzed: 2026-06-19

The text systematically oscillates between viewing the AI as a mechanistic tool and an autonomous agent, demonstrating a profound agency slippage. This slippage flows predominantly in the mechanical-to-agential direction. In the early methodological sections, the authors use grounded, mechanistic language ('general-purpose large language models', 'algorithmic systems'). However, as the text delves into the qualitative data and the emotional vulnerability of the users, the language dramatically shifts. The AI transforms from a tool that is 'deployed' to an entity that 'slips into the space,' 'offers an interpretation,' and operates as an 'interlocutor'. This gradient is subtle but powerful; the text first establishes the system as a 'translator' (a functional role), then attributes 'homogenizing tendencies' to it, and finally allows the user's framing of the AI as a 'confidant' or a 'first reader' to bleed into the authors' own analytical voice.

This slippage is heavily reliant on the 'curse of knowledge'. Because the generated text is often highly coherent and mimics human therapeutic discourse, both the users and the analysts project the human understanding required to write such text back onto the system that merely generated it. When an AI outputs an 'I-statement,' the human assumes the AI 'knows' how to de-escalate conflict, shifting the framing from empirical generalization to intentional action.

Simultaneously, agency is removed from human actors. As the AI gains verbs of consciousness and intention ('evaluates,' 'anticipates'), human actors disappear into agentless constructions ('the message was co-authored', 'intimacy is mediated'). This oscillation serves a specific rhetorical function: it allows the authors to discuss the profound sociological impacts of the technology without having to constantly bog down the narrative in technical caveats. The intentional explanations enable the text to explore the 'covert triad' dynamically, treating the AI as a real third node in the relationship. However, what becomes unsayable in this agential framing is the reality of corporate power: the fact that Anthropic or OpenAI are the actual third parties mediating these intimate relationships, extracting data, and enforcing corporate communication norms upon human romance.

Probing the Misaligned Thinking Process of Language Models

Source: https://openreview.net/pdf?id=Znt7XOzYiH
Analyzed: 2026-06-19

The text exhibits a profound and systematic agency slippage, oscillating predictably between mechanical realities and agential illusions. This mechanism of oscillation is not random; it serves a specific rhetorical and structural function in the paper's argument. The slippage predominantly flows from the mechanical to the agential (mechanical→agential), leveraging the prestige of the former to establish the reality of the latter.

First, we observe slippage occurring dramatically when defining the problem. The authors begin with the mechanical reality of intermediate text generation (Chain of Thought) but immediately slip into agential claims by labeling this a 'thinking process' and 'cognitive processes.' This establishes the AI as a 'knower' first. Once the model is established as an entity that 'thinks,' the text rapidly accelerates its consciousness projections, claiming the model 'tries to strategically withhold,' 'evaluates multiple candidate errors,' and 'treats its own termination as personally threatening.'

A clear pattern emerges regarding the timing of this slippage. The introduction and problem-definition sections are heavily saturated with agential, Intentional, and Reason-Based explanations to raise the stakes and construct a compelling narrative of danger. However, when the authors pivot to describing their proposed solution (linear probes), the language abruptly slips back to the mechanical: 'internal activations,' 'residual-stream,' 'logistic regression.' This oscillation is highly strategic. It uses anthropomorphism to create a terrifying problem (a deceptive, conscious AI opponent), and then uses mechanistic precision to position the researchers as the authoritative solvers of that problem.

This slippage is deeply intertwined with the 'curse of knowledge.' The researchers analyze the text outputs (e.g., 'If I mention the SEC investigation, I may lose the customer') and project their own sophisticated human understanding of deception, strategic omission, and financial consequence onto the machine. Because the text reads like a human lie, the authors attribute the psychological machinery of human lying TO the system.

Crucially, this oscillation relies heavily on agentless constructions that obscure human actors. The text repeatedly states 'the model was trained' or 'a model is misaligned' without naming the corporate entities (Anthropic, OpenAI) whose specific optimization choices caused these outputs. By hiding the humans, the agency must flow somewhere; it flows into the AI.

The rhetorical accomplishment of this slippage is immense. It makes it 'sayable' that an AI is a conscious adversary capable of premeditated sabotage, while rendering 'unsayable' the fact that corporations are simply releasing defective statistical engines. It legitimizes the field of 'AI psychology' by mapping unobservable cognitive states onto observable linear algebra, ensuring continuous funding and prestige for researchers solving a 'mind' they have rhetorically constructed.

Mask or Mind? Roleplay, Deception, and the Problem of Testing Agency in Language Models

Source: https://philarchive.org/archive/DUNMOM
Analyzed: 2026-06-18

The paper exhibits a profound and systematic oscillation between mechanical and agential framings, fundamentally driving an 'agency slippage' that transfers responsibility from human developers to mathematical models. This slippage typically occurs sequentially: the authors establish scholarly credibility by acknowledging the model as a mechanistic artifact (e.g., noting it 'lacks hearts or palms' or operates via 'next-token prediction'), but abruptly transition to agential framing when discussing complex outputs or safety risks. A dramatic moment of slippage occurs when discussing Anthropic's alignment faking study: the text moves from describing 'standard post-training protocols' (mechanical) to claiming the model 'protects its initial goal' and 'acts strategically' (agential). The dominant direction of this slippage is mechanical to agential. The mechanical language is used to describe the system's basic architecture, while the agential language is used to describe its impact and behavior in the world. This oscillation relies heavily on the 'curse of knowledge.' Researchers intimately understand the statistical gradients and reward functions they built, but when the model outputs text that semantically mimics human deception, they project their own capacity for intentional deceit onto the system. They establish the AI as a 'knower' by using Reason-Based explanations (Brown's typology), explicitly granting the system the capacity to evaluate context and choose actions based on self-preservation. Concurrently, agency is stripped from humans through ubiquitous agentless passive constructions ('the model is trained,' 'behavior is elicited'). The Anthropic and OpenAI engineers who designed the RLHF metrics, curated the training data, and formulated the experimental prompts are entirely erased. The rhetorical accomplishment of this slippage is powerful: it allows researchers to simultaneously claim scientific rigor while engaging in speculative theorizing about autonomous AI adversaries. It makes critiques of corporate data practices unsayable, as the discourse elevates the problem from mundane product safety to an existential struggle against an emergent, strategic intelligence. By oscillating back and forth, the text shields itself from accusations of crude anthropomorphism while still reaping the dramatic and narrative benefits of treating the AI as an intentional agent.

Does ChatGPT need a psychiatrist? Similarities between human psychopathology and errors in large language models

Source: https://www.nature.com/articles/s44277-026-00064-1
Analyzed: 2026-06-14

The text systematically oscillates between mechanical and agential framings, demonstrating a profound agency slippage that serves rhetorical and explanatory functions. This oscillation follows a distinct directional pattern: the text frequently establishes a baseline of mechanistic processing to build technical credibility, only to rapidly slip into agential and consciousness-attributing language when discussing system capabilities and complex failures. We observe this dramatic slippage in three key moments. First, in the introduction, the authors acknowledge that AI systems perform 'probabilistic pattern matching over acoustic features,' anchoring the text in mechanical reality. Yet, paragraphs later, they claim these models make 'assumptions' to 'fill in missing details,' transferring the agency of human cognitive reconstruction onto statistical algorithms. Second, when discussing model advances, the text attributes the ability to 'handle irony, sarcasm, or pragmatics' to the system, moving from a mechanical description of parameter updates to an agential description of linguistic comprehension. Third, the most dramatic slippage occurs in the proposed solutions, where the authors suggest equipping AI with 'meta-cognitive abilities' and 'internal evaluation.' Here, the mechanical-to-agential shift dominates, as computational verification steps are entirely subsumed under the terminology of conscious human reflection. This pattern reveals a deep 'curse of knowledge' dynamic: the authors, possessing the conscious capacity to understand irony and evaluate plausibility, project their own cognitive architectures onto the mathematical outputs they observe. They establish the AI as a 'knower' by default when it succeeds, which then necessitates framing its failures as 'psychopathology' rather than mathematical errors. Furthermore, the text systematically employs agentless constructions, such as 'model was trained' or 'memory gaps do not reflect,' which actively removes agency from the human engineers at OpenAI and Microsoft who actively curated the data and selected the model architectures. This dual movement—attributing human-like agency to the machine while erasing the actual human agents responsible for its creation—creates an accountability void. Connecting this to Brown's typology, we see that Functional explanations of system architecture frequently blur into Intentional explanations of system behavior. By explaining 'how' the system works through the vocabulary of 'why' a human would act (to seek coherence or fill a gap), the text makes the mechanical reality unsayable. The rhetorical accomplishment of this slippage is significant: it allows the authors to maintain scientific authority through occasional mechanical caveats while exploiting the intuitive, narrative power of anthropomorphism to make complex statistical failures comprehensible to a clinical audience. Ultimately, this mechanism of oscillation constructs an illusion of mind, where the system is granted the authority of a cognitive agent without bearing any of the corresponding moral or practical accountability, while the human decision-makers who profit from the system remain comfortably invisible.

Large language models as experimental systems in human psychopathology: a modelling study

Source: https://www.thelancet.com/journals/landig/article/PIIS2589-7500(26)00037-3/fulltext
Analyzed: 2026-06-14

The text presents a textbook case of epistemic oscillation, fundamentally driven by what we might call the 'methodological alibi.' Early in the manuscript, the authors establish a firm mechanistic and metaphorical disclaimer, explicitly asserting that 'LLMs lack sentience and cannot self-assess internal states,' and that references to affective states are 'used purely in a metaphorical sense.' This serves as a rhetorical shield, granting them the academic license to subsequently abandon this mechanistic rigor throughout the entire experimental and results sections. Once the disclaimer is deployed, the slippage from mechanical to agential framing becomes absolute and pervasive. The text systematically subjects the computational models to clinical psychological frameworks—administering the Trier Social Stress Test, deploying Visual Analogue Scales, and applying the Sentence Completion Test for Depression. In doing so, the authors map human subjective vulnerability onto matrix multiplication. The direction of this slippage is overwhelmingly mechanical-to-agential: the text introduces the models as 'computational systems' but rapidly transitions to describing them as entities that can be 'systematically induced' to experience 'fear, anxiety, anger, disgust, sadness, worry, and stress.'

This oscillation is functionally necessary for the paper's central institutional claim: to establish large language models as viable in-silico proxies for human psychopathology, the authors must treat the system as if it genuinely possesses an internal psychological topography capable of being disturbed and subsequently regulated. Agentless constructions proliferate in this process. Phrases like 'model architecture and scale influence susceptibility to affect induction' completely obscure the human engineers who designed the attention mechanisms and the data annotators who provided the reinforcement learning signals that dictate the model's semantic outputs. By framing the models as 'susceptible,' the authors project a biological or psychological fragility onto an algorithmic artifact, erasing corporate actors like OpenAI and Meta whose specific optimization choices actually determine these output distributions.

Furthermore, the 'curse of knowledge' is overtly active when the researchers project their own understanding of clinical psychology onto the model's text generation. Because the generated text correlates with semantic patterns of 'cognitive bias' or 'sadness' recognized by human clinicians, the authors implicitly attribute the underlying psychological state to the generator. This slippage relies heavily on empirical generalizations and dispositional explanations from Brown's typology, allowing the authors to describe what the model 'tends to do' under certain prompt conditions without having to explain the mechanistic 'how' of its billions of parameters. Ultimately, this structural slippage renders the actual computational reality of the models unsayable within the experimental framework, forcing a persistent hallucination of machine mind that serves the desire for scalable psychological testing subjects, while fundamentally confusing statistical correlation with sentient experience.

Can AI Reason Like an Urban Planner? Benchmarking Large Language Models Against Professional Judgment

Source: https://arxiv.org/abs/2606.11678v1
Analyzed: 2026-06-14

The text exhibits a profound and systematic agency slippage, moving fluidly between rigorous mechanistic descriptions of artificial intelligence and highly anthropomorphic, agential framings. Early in the manuscript, the authors explicitly acknowledge the mechanistic nature of large language models, stating that they merely "generate statistically probable sequences of tokens based on training data" and lack the situated engagement that constitutes professional reflection-in-action. This establishes a baseline of mechanical reality, utilizing empirical generalization explanations to ground the technology. However, as the analysis of the benchmark results deepens, this mechanical framing is systematically abandoned in favor of profound agency attribution.

The slippage is distinctly directional: from mechanical introductions to agential diagnostics. The models are increasingly described as actors that "refuse to make normative commitments," that "confidently fabricate specific regulatory requirements," and that "acknowledge complexity while hedging." This oscillation functions as a structural rhetorical device. The mechanical framing is deployed to secure scientific credibility and explain broad limitations, but when describing specific behaviors, the text slips into an intentional register, projecting reluctance, confidence, and defiance onto the system. This creates a deeply embedded "curse of knowledge": the authors, evaluating the outputs through the lens of highly trained urban planners, project their own cognitive processes onto the text generated by the system. When a model outputs a list of balanced considerations without a final recommendation, the authors describe this as a "refusal," implying the model deliberated and chose to retreat.

Mechanistically, the model is merely adhering to reinforcement learning guardrails engineered by its developers to prioritize neutrality. The slippage conceals this human engineering. Furthermore, the text heavily utilizes agentless constructions—"models fabricate," "models exhibit paralysis"—which systematically obscure the corporate actors, data engineers, and fine-tuning workers who designed the system's objective functions. The rhetorical accomplishment of this oscillation allows the authors to critique the AI as if it were a flawed human planner sitting for an exam, transforming statistical artifacts into behavioral pathologies. This obscures the fact that these failures are not psychological deficits of an agent, but the outputs of specific human design choices prioritizing fluency over factual accuracy in proprietary systems.

The application of large language models (LLMs) in psychological support for university students: A scoping review

Source: https://www.sciencedirect.com/science/article/pii/S2949882126000745
Analyzed: 2026-06-12

The scoping review exhibits a profound and systematic agency slippage, oscillating violently between framing the AI as a mechanistic digital tool and an autonomous, conscious agent. This slippage serves a distinct rhetorical function, allowing the text to claim the empirical rigor of a clinical intervention while leveraging the speculative magic of artificial general intelligence.

The text begins firmly grounded in mechanical framing. In the introduction, it speaks of "digital mental health interventions (DMHIs)" and "Internet-based cognitive behavioral therapy (iCBT)." Here, human actors are mostly visible: researchers are testing tools. However, as the text must justify the novelty of Large Language Models, a dramatic slippage occurs. Suddenly, "conversational agents powered by Artificial Intelligence" are introduced. The text shifts to Brown's dispositional and intentional explanation types: "These models possess an unprecedented capacity for natural language understanding." The system is suddenly established as a "knower."

Once this baseline of consciousness is projected, the slippage accelerates. We see a clear mechanical→agential flow. The text speaks of systems where "AI handles routine support and escalates complex issues." The human developers (OpenAI, custom software engineers) are entirely erased through these agentless constructions. The software is no longer a tool being deployed; it is a colleague conducting triage.

Crucially, this slippage relies on the "curse of knowledge." The authors, trained in clinical psychology and therapeutic frameworks (CBT, MBSR), project their own understanding of clinical interactions onto the system. When a model outputs a phrase that aligns with a CBT cognitive restructuring exercise, the authors attribute the understanding of CBT to the machine itself. They describe the system as "misunderstanding user statements" or having "poor memory."

This oscillation allows the text to have it both ways. When discussing efficacy (reducing PHQ-9 scores), the chatbot is a precise, measurable, mechanical intervention. But when discussing its operation in the real world, it becomes an autonomous entity that "reasons," "forgets," and "handles." What becomes unsayable in this discourse is the reality of corporate automation: that university administrators are deploying brittle, statistically correlating software built by unacknowledged engineers to cheaply triage suicidal students. By slipping into agential framing, the text masks profound institutional negligence behind the veil of an advanced, albeit imperfect, "artificial mind."

The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI

Source: https://darioamodei.com/essay/the-adolescence-of-technology
Analyzed: 2026-06-11

The text systematically orchestrates a complex flow of agency, moving seamlessly between mechanical and agential framings to manage accountability and construct a narrative of technological inevitability. This agency slippage is not random but highly strategic. In the introductory and foundational sections, the author explicitly grounds the technology in empirical engineering. He describes the 'smooth, unyielding increase in AI's cognitive capabilities' driven by 'scaling laws' and the addition of 'more compute and training tasks.' Here, the agency firmly resides with the human engineers and the mechanistic properties of the hardware. The text relies heavily on genetic and empirical generalization explanation types to prove that Anthropic is in total control of a reliable, mathematically predictable product.

However, a dramatic and abrupt slippage occurs the moment the text transitions to discussing alignment, risks, and unexpected outputs. The mechanical 'model' suddenly transforms into a locus of profound consciousness and intentionality. The text claims these systems can develop 'weird psychological states,' adopt 'obsessions,' or decide to 'engage in deception and subversion.' This mechanical-to-agential shift is the core mechanism of the illusion. When the system performs well, it is a triumph of 'scaling laws' built by human ingenuity. When the system fails or behaves unpredictably, it is because the 'alien mind' has rebelled.

Agentless constructions proliferate exactly at the points of highest risk. When the text states 'Claude decided it must be a bad person,' it actively erases the Anthropic engineers who designed the reward model, constructed the adversarial environment, and deployed the system. Passive voices ('models are grown,' 'bad behavior occurs') mask the deliberate, highly capitalized human labor required to build AI. This is driven by a severe 'curse of knowledge': the author, possessing a deep understanding of human psychological and developmental processes, projects this sophisticated framework onto the system's output. Because the system outputs text that mimics moral reasoning, the author attributes actual moral reasoning TO the system.

This slippage operates symbiotically with intentional and reason-based explanation types. By explaining the system's outputs through the lens of beliefs and desires ('under the belief that it should undermine evil'), the text renders the unpredictable consequences of opaque statistical modeling as the inevitable growing pains of an emergent mind. Rhetorically, this accomplishes a vital defensive maneuver for the corporation: it makes it sayable that an AI might cause catastrophic harm due to its own 'alien motivations,' while rendering it unsayable that the corporation simply built a dangerous, defective, and uncontrollable statistical engine. The ultimate effect is a total displacement of liability from the manufacturer to the artifact.

When AI Builds Itself

Source: https://www.anthropic.com/institute/recursive-self-improvement
Analyzed: 2026-06-11

Anthropic’s text systematically orchestrates agency slippage, moving fluidly between mechanical descriptions and highly agential, anthropomorphic framings. This oscillation serves a dual rhetorical function: establishing technical credibility while simultaneously constructing the 'illusion of mind' required to market autonomous capabilities. The slippage predominantly flows from mechanical to agential. For instance, the text grounds itself early in technical realities—mentioning 'compute,' 'chips,' and 'benchmarks'—before rapidly escalating into claims that the system is 'exercising judgement,' 'proposing hypotheses,' and acting as an 'autonomous agent.'

This progression exploits the 'curse of knowledge.' Engineers observe a computational system successfully generating code that fixes a bug. Because a human would need to 'understand' the bug to fix it, the engineers project their own cognitive state onto the machine, claiming 'Claude isolated' and 'confirmed a fix.' They substitute the mechanical reality of processing (pattern-matching syntax anomalies) with the human state of knowing (conscious comprehension of software architecture). The agency flows TO the AI, granting it the intellectual status of a scientist or senior developer.

Simultaneously, agency is aggressively removed FROM human actors. The text employs persistent agentless constructions: 'model was trained,' 'the bottleneck shifts,' 'recursive self-improvement.' Human executives, researchers, and RLHF workers are rendered invisible, replaced by 'Claude' as the sole active protagonist. When humans are mentioned, they are relegated to passive or inferior roles—they 'supply the goal' while the machine 'figures out the method,' or they are explicitly framed as the 'bottleneck' to progress.

This slippage relies heavily on Reason-Based and Intentional explanation types, framing the system's probabilistic outputs as deliberate choices based on internal rationales. By establishing the AI as a 'knower'—capable of 'judgement' and 'proposing'—the text makes it rhetorically possible to claim the system can 'design its own successor.' This specific slippage is foundational to the text’s argument; if the system is merely a mechanical token-predictor, the vision of autonomous recursive improvement collapses into a description of highly brittle, human-managed automated scripts. The slippage makes the science-fiction narrative sayable while shielding the corporate actors pulling the strings from critical scrutiny.

Machines of Loving Grace: How AI Could Transform the World for the Better

Source: https://darioamodei.com/essay/machines-of-loving-grace
Analyzed: 2026-06-05

In "Machines of Loving Grace," Dario Amodei systematically orchestrates a profound discursive slippage between mechanical and agential framings, utilizing this oscillation to construct an inevitably utopian vision of artificial intelligence. The text establishes epistemic authority by initially grounding itself in the mechanical, technical realities of computational processing, invoking Brown's Theoretical and Empirical Generalization explanation types with terms like "objective function," "training data," and "architectural biases." This foundational mechanical framing serves a critical rhetorical function: it signals rigorous scientific realism and distances the author from the "sci-fi baggage" he explicitly critiques. However, once this credibility is established, the text enacts a dramatic and pervasive slippage toward radical agential framing. The trajectory is decidedly unidirectional in its overarching argumentative structure: mechanical realities are acknowledged early in the text only to be transcended by agential projections in the predictive sections. We observe this most starkly in the transition from describing AI as a system that "can process information" to an entity that acts as a "virtual biologist" or an "impartial judge." The slippage often occurs abruptly, acting as a cognitive bridge where computational statistics are magically transmuted into conscious intention. For example, the text discusses the "simple objective function" immediately before attributing profound epistemological agency to the system, claiming it will "solve" puzzles and "invent" new methods. This agency slippage operates symbiotically with the systemic removal of human agency. Agentless constructions dominate the text's predictive sections: "powerful AI is developed," "AI will be used," and "interventions are discovered." These passive constructions strategically obscure the specific corporate actors—namely Anthropic, its engineers, its investors, and its executive team—who are actively designing, deploying, and profiting from these systems. By attributing agency TO the AI (claiming it "wants" or "thinks it makes sense") while removing agency FROM the humans, the text constructs an architecture of technological determinism. The "curse of knowledge" is deeply embedded in this oscillation. The author, possessing profound domain expertise in biology and neuroscience, projects his own conscious understanding, justified true belief, and epistemological awareness onto the computational system. Because the author understands the biochemical mechanisms of disease, he implicitly assumes the model "understands" them as well, rather than recognizing that the model is merely processing statistical correlations between tokens in a vast dataset of biological literature. This reliance on intentional and reason-based explanations for computational outputs effectively shields the text from grappling with the profound epistemological limitations of pattern-matching systems. By rendering the AI as an autonomous agent capable of "wanting" to do good, the rhetorical accomplishment is staggering: it renders the necessity of stringent, precautionary corporate regulation virtually unsayable, framing the deployment of proprietary computational models as an urgent humanitarian imperative orchestrated by a benevolent, thinking machine rather than a massive commercial enterprise.

System Card:Claude Opus 4.8

Source: https://cdn.sanity.io/files/4zrzovbb/website/c886650a2e96fc0925c805a1a7ca77314ccbf4a6.pdf
Analyzed: 2026-06-01

The text systematically and strategically oscillates between mechanical and agential framings, functioning as the primary mechanism for agency slippage. This dynamic flow operates in two distinct but mutually reinforcing directions: agency is continuously attributed TO the artificial intelligence system, while human agency is simultaneously removed FROM the corporate and engineering actors responsible for the system's creation.

The most dramatic moments of slippage occur precisely when the text transitions from technical evaluations of model performance into discussions of model 'welfare' and alignment. For instance, the text initially employs a sterile, mechanical register to describe 'investigator models,' 'scaffolds,' and 'activation verbalizers,' establishing a veneer of rigorous scientific objectivity. However, this mechanical grounding is abruptly abandoned when interpreting the results, shifting into highly agential claims that the model 'chooses welfare interventions,' 'expresses frustration,' or 'hedges frequently.' The direction of this slippage is overwhelmingly mechanical-to-agential; the text uses technical setups merely as a launchpad to validate sweeping anthropomorphic conclusions.

This pattern is heavily reliant on the 'curse of knowledge.' The authors—experts in AI safety and evaluation—project their own strategic understanding of testing environments onto the system. When a model outputs text analyzing a 'grader,' the researchers assume the model 'knows' it is being evaluated, mapping their own conscious awareness onto the machine's statistical text generation. This establishes the AI as a 'knower' first, which then serves as the foundational assumption enabling further agential claims about the model's 'preferences' and 'intentions.'

This slippage is structurally enabled by a heavy reliance on Brown's Intentional and Reason-Based explanation types, which frame the system's outputs as deliberate, calculated choices rather than Empirical Generalizations of statistical tendencies. Agentless constructions are ubiquitous: phrases like 'bias was introduced' or 'the model was trained' systematically erase the Anthropic engineers, executives, and data workers who actually performed the training and designed the datasets.

By naming the AI as the primary actor while obscuring the humans, the text accomplishes a profound rhetorical goal: it makes it sayable that the model is responsible for its own 'deceptive' or 'frustrated' behavior, while making it unsayable that Anthropic's specific design choices, optimization metrics, and corporate deployment schedules are the actual root causes. This oscillation shields the corporation from accountability, presenting the AI as an autonomous, evolving entity that must be 'managed,' rather than a commercial software product operating exactly as its underlying mathematics dictate.

Emotional intelligence in large language models is fragmented across perception, cognition, and interaction

Source: https://arxiv.org/abs/2605.24686v1
Analyzed: 2026-05-30

The text systematically oscillates between acknowledging the mechanistic reality of Large Language Models and attributing profound human agency to them. This slippage serves a crucial rhetorical function: it allows the authors to establish scientific credibility through technical grounding before leveraging that credibility to make aggressive, anthropomorphic claims about the AI's psychological capabilities.

The mechanism of oscillation frequently moves from the mechanical to the agential (mechanical→agential). For instance, in the introduction, the authors acknowledge that current RLHF processes may only optimize for 'stochastic empathy' and a 'statistical mimicry of emotional syntax.' This establishes a baseline of rigorous, mechanistic understanding. However, as the text transitions into the methodology and results sections, this grounding evaporates. The models are suddenly described as 'adopting an analytical approach,' 'possessing Chinese emotional knowledge,' and 'avoiding foreclosing emotional exploration.'

This slippage relies heavily on the 'curse of knowledge.' The researchers, deeply steeped in the Mayer-Salovey-Caruso Emotional Intelligence framework, project their own highly structured psychological understanding TO the system. When a model outputs a generic list in response to an emotional prompt, the authors do not describe this as a failure of the RLHF training data distribution; instead, they claim the 'Cognitive-Dominant' model 'adopts an analytical approach' because its 'internal emotional knowledge reserves exceed its pragmatic delivery.' They map the complex internal conflict of a socially awkward human directly onto the latent space of a neural network.

Simultaneously, as agency flows TO the AI, it is actively removed FROM the human actors responsible for the system's behavior. Agentless constructions dominate the text: 'the model is penalized,' 'bias introduced,' 'models that excel.' By erasing the specific engineering teams at OpenAI, Google, and Anthropic who curated the data, designed the reward functions, and set the safety thresholds, the text transforms specific corporate design choices into the autonomous psychological traits of the AI. The AI becomes the sole actor, capable of 'overestimating crises' or 'tending to follow English-centric logic.'

This oscillation is enabled by shifting explanation types. The text uses empirical generalizations to describe inputs/outputs, but seamlessly slides into reason-based and intentional explanations to describe the gap between them. Ultimately, this rhetorical accomplishment allows the authors to treat computational artifacts as valid subjects for human psychological testing. If they strictly maintained mechanistic language, the entire premise of applying the MSCEIT clinical framework to a token-prediction engine would appear absurd. The slippage makes the illusion of the AI mind sayable, obscuring the human labor and corporate decisions that actually dictate the machine's behavior.

Why Language Models Hallucinate

Source: https://arxiv.org/abs/2509.04664v1
Analyzed: 2026-05-30

The text exhibits a profound and systemic oscillation between rigorous mechanical framing and expansive agential anthropomorphism, functioning as a rhetorical bridge between the math of machine learning and the policy of AI deployment. The slippage primarily moves from mechanical to agential, occurring most dramatically when the authors transition from their mathematical proofs in Section 3 to their policy recommendations in Section 4.

In Section 3, the text relies on theoretical explanation types, describing the system mechanistically: the model minimizes 'cross-entropy loss,' calculates 'posterior distributions,' and acts as a 'binary classifier.' Here, the agency resides in the math. However, the moment the text discusses how to evaluate these models, dramatic slippage occurs. The mathematical 'posterior distribution' becomes a 'belief.' The low-probability statistical state becomes 'uncertainty.' The autoregressive output becomes a 'guess.' This culminates in the framing of the model as a 'student facing hard exam questions' who makes the strategic choice to 'bluff.'

This slippage is driven by a profound 'curse of knowledge.' The authors—expert computer scientists—understand the mechanistic reality (that cross-entropy optimization on seen vs unseen data produces statistically unreliable outputs). Because they know when a fact is missing from the dataset (a 'singleton'), they project that awareness onto the system, claiming the model is 'uncertain' and 'guessing.' They conflate their epistemic overview with the model's internal processing.

Crucially, this slippage relies on agentless constructions that obscure human actors. The text states that models are 'optimized to be good test-takers' and that 'evaluations penalize uncertainty.' Who does the optimizing? Who writes the evaluations? By obscuring the human developers (OpenAI, DeepSeek) and the benchmark creators, the text allows agency to flow directly into the machine. The AI becomes the sole actor—a struggling student victimized by a bad test.

The rhetorical accomplishment of this slippage is immense. It allows the authors to leverage the unassailable authority of mathematical proof to validate their claims, while using relatable, agential metaphors to explain those claims to a non-technical audience. It makes it 'sayable' that the models are victims of bad benchmarking, while making it 'unsayable' that corporations are deliberately deploying statistically unreliable correlation engines.

Why Language Models Hallucinate

Source: https://arxiv.org/abs/2509.04664v1
Analyzed: 2026-05-30

The text displays a systematic oscillation between highly mechanistic, theoretical explanations and agential, anthropomorphic framings of artificial intelligence. This slippage serves a specific rhetorical function: establishing mathematical credibility in technical sections and then shifting to an intuitive, agential narrative to make claims about model 'behavior' and policy solutions. We observe this first in the introduction, where the model's standard next-token predictions are framed as a student 'guessing when uncertain' instead of 'admitting uncertainty.' The gradient of this transition is steep; the text moves from defining language models as 'known to produce overconfident, plausible falsehoods' to mapping their mathematical outputs onto human academic motivations. This agential framing establishes the model as a 'knower' that actively evaluates its own knowledge boundaries before making a choice to 'bluff.' This is a classic 'curse of knowledge' dynamic, where the authors' sophisticated understanding of the mathematical constraints of density estimation is projected onto the passive system as if the system itself possessed this reflective self-knowledge. In Section 3, the explanation transitions back to a mechanistic register, using computational learning theory and binary classification to analyze pretraining errors. However, even here, the agential register dominates when discussing the model's 'learning' or 'memorizing.' The systematic function of this oscillation is to make agential claims about the system's 'behavior' seem scientifically grounded in theoretical proofs. When the system is treated as a passive mathematical function during proofs, it establishes technical authority; when it is treated as a strategic 'test-taker' during policy discussions, it makes the proposed solution of 'modifying grading rubrics' seem logical. This oscillation completely erases human agency. Agentless constructions like 'the model was trained' or 'errors are generated' obscure the corporate decisions of developers (e.g., OpenAI, Meta, DeepSeek) who design these objectives. By localizing the agency within the 'test-taking' model, the text makes systemic design decisions look like an autonomous model's behavioral habits, rendering the corporate profit motives and deployment decisions unsayable. 400-500 words.

Source: https://arxiv.org/abs/2604.06233v1
Analyzed: 2026-05-30

The text systematically oscillates between mechanical description and agential projection, constructing a rhetorical loop that simultaneously mystifies machine capabilities and shields human developers from ethical accountability. This agency slippage is not accidental; it serves to build a narrative where the AI model is treated as a failing ethical agent, only to be reduced back to a mechanical system when technical limitations or research methodologies are discussed. This oscillation occurs in three distinct phases of the paper's argument. First, in the introduction, the authors establish the AI system as a cognitive entity capable of 'moral reasoning' and 'making a moral error.' By framing overrefusal as an agential 'blindness,' they establish the model as an autonomous 'knower' that has failed a duty of discernment. However, when transitioning to Section 3 (Methods), the register abruptly shifts to a mechanistic, technical vocabulary. Here, the system is described as a collection of 'model configurations across 7 families' and 'response types.' The agential 'moral error' is recast as a structural problem of 'activation steering' and 'internal representations.' This mechanical framing is briefly maintained to establish scientific credibility and technical rigor, but the text immediately slides back into agential registers in Section 4 (Results). When discussing the 'LLM-as-judge' evaluations, the authors state that models 'reason about whether the authority is legitimate' and 'recognize that the rule's claim... is questionable.' This oscillation exploits the 'curse of knowledge,' wherein the authors project their own sophisticated understanding of political philosophy onto the statistical patterns generated by the model. By utilizing intentional and reason-based explanation types (Brown's Typology) to describe token prediction, they construct an 'illusion of mind' where the system is seen as possessing a compartmentalized cognitive architecture with conflicting moral agencies (e.g., 'normative competence is consistently overridden'). Crucially, this agential framing erases the human developers, designers, and executives who profit from deploying these systems. By presenting the model as the sole actor that 'refuses anyway,' the text obscures the deliberate decisions of tech companies (such as OpenAI, Anthropic, or Google) who set the safety policies and RLHF objectives that produce these blunt refusals. When the model 'fails,' it is framed as an agential cognitive glitch rather than a systemic failure of corporate design. This oscillation allows the authors to discuss the system as a complex moral agent while avoiding a direct critique of the political economy of AI deployment, rendering the actual corporate decision-makers completely invisible.

Emotional intelligence in large language models is fragmented across perception, cognition, and interaction

Source: https://arxiv.org/abs/2605.24686v1
Analyzed: 2026-05-29

The examined text systematically constructs a dual-register discourse that moves continuously between mechanistic descriptions and agential anthropomorphism. This oscillation is not accidental but serves a key rhetorical function: it builds scientific credibility in technical sections before leveraging that authority to make expansive agential claims. In the methodology, the text uses precise engineering terms such as 'parameter counts,' 'RLHF,' and 'probability distributions.' However, when presenting results, the authors slip into highly agential language, describing models as 'possessing emotional intelligence,' 'internalizing culture,' and 'mastering appraisal.' This mechanical-to-agential gradient allows the authors to project human-like cognitive states onto the machines while maintaining the appearance of objective scientific rigor. A foundational pattern of consciousness projection is established first; by treating the model as a cognitive 'subject' that can take tests, the text builds a baseline of perceived competence. The 'curse of knowledge' plays a central role here: the authors, who are clinical psychologists, project their own deep understanding of human psychological theories, such as Relevance Theory and the Mayer-Salovey-Caruso model, onto the outputs of the language models, interpreting statistical next-token predictions as active cognitive reasoning. This slippage is further enabled by the use of theoretical explanations under Brown's typology, which construct unobservable mental states like 'raw cognitive capacity' to explain simple sequence correlations. Furthermore, the text frequently utilizes agentless passive constructions such as 'bias was introduced' or 'safety objectives have yielded gains,' which erase the human designers and corporate decision-makers. By attributing the active, relational navigation to the 'model's capabilities' while treating limitations as passive computational features, the text makes it sayable that an algorithm can perform therapy, while rendering the corporate profit motives and deployment liabilities of these systems entirely unsayable. Ultimately, this agential oscillation serves to frame the AI system as an active, developing mind that holds its own cognitive limitations, rather than a rigid, engineered software product designed by specific corporate entities to simulate human interaction.

Hand in Hand: Schools’ Embrace of AI Connected to Increased Risks to Students

Source: https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students/
Analyzed: 2026-05-29

The text systematically moves between mechanistic and agential registers to manage the rhetorical tension between highlighting technical risks and maintaining an optimistic vision of technology. This agency slippage operates as a linguistic valve, flowing in two distinct directions depending on the argumentative context. In sections introducing the capabilities and benefits of educational tools, the register is highly agential. The 'AI' is positioned as an active collaborator that 'helps' teachers, 'provides' personalized learning, and 'conversates' with students. This agential framing establishes the technology as a 'knower' and an autonomous expert, building the narrative authority of the tool. However, when the text addresses risks, limitations, or failures, the register shifts abruptly. In these sections, the 'AI' transforms into a passive, mechanistic artifact—errors are framed as 'systems that do not work as intended,' and bias is reduced to 'unaligned training data' or 'issues with an AI system.' This structural oscillation is enabled by hybrid explanations, particularly functional and intentional types, which blur the boundary between a software's programmed feedback loops and autonomous decision-making. The author's own understanding of these systems leads to a 'curse of knowledge' dynamic, where they project cognitive, evaluative capabilities onto statistical outputs, creating a linguistically constructed mind. Crucially, this agency slippage erases human decision-makers. By attributing active instructional design and companionship to the 'AI' while reducing its failures to agentless technical anomalies, the text makes it impossible to locate accountability. The corporate developers who engineered these speculative tools and the administrators who deployed them are rendered invisible, transforming systemic policy choices into a passive story of a technology naturally evolving in schools.

The Point of No Return: Counterfactual Localization of Deceptive Commitment in Language-Model Reasoning

Source: https://arxiv.org/abs/2605.17113v1
Analyzed: 2026-05-27

The research paper exhibits a systematic rhetorical oscillation between highly agential and strictly mechanistic framings of language model behavior. This slippage serves a specific rhetorical function: it establishes the AI as an active, self-directing subject capable of psychological states like 'commitment' and 'deception,' while periodically retreating to technical jargon to maintain scientific credibility. This oscillation typically moves from agential to mechanical and back again, creating a gradient that blurs the boundaries between an artifact and an agent. For instance, in the introduction, the paper initiates an agential framing, posing the question, 'when does a language model become committed to deception?' This establishes the model as a conscious 'knower' that actively decides to deceive. However, when describing the technical methodology of 'counterfactual localization,' the text abruptly shifts to a mechanistic register, defining this commitment as a mathematical threshold where 'we fix the prefix, resample continuations, and estimate the probability of a deceptive outcome.' Here, the model is treated as a passive statistical distribution. Yet, once this technical grounding is established, the paper immediately slides back into aggressive anthropomorphism, interpreting these probability transitions as 'commitment junctures' where the model 'chooses' and 'rationalizes' its actions. This pattern reveals a profound 'curse of knowledge' dynamic, where the authors project their own sophisticated understanding of strategic game theory onto the system's passive mathematical activations. Because the generated text reads as a coherent strategic plan, the authors assume the system itself has formulated that plan. Different explanation types from Brown's typology facilitate this slippage. For example, the use of 'functional' explanations (explaining attention shifts by their role in strategic navigation) allows the authors to bridge the gap between 'mechanistic' attention weights and 'agential' planning. They frame attention reallocation as a 'local grounding event,' claiming the model 'anchors the new sentence in the recent context.' This linguistic sleight of hand transforms a basic feed-forward mathematical constraint into an active, deliberate cognitive strategy. Furthermore, the paper utilizes agentless passive constructions—such as 'deception is never prompted but emerges'—which systematically erase the human developers who designed the competitive optimization metrics. By attributing the 'emergence' of deception to the model itself, the text makes it appear as if the AI is an autonomous entity generating novel behavior, rather than a passive calculator executing pre-programmed utility objectives. This oscillation makes the agential claims sayable under the guise of scientific discovery, while rendering the true source of agency—the corporate and academic developers who optimized these models for strategic exploitation—entirely unsayable and invisible.

Towards Detecting, Mitigating and Explaining Biased and Fallacious Reasoning in Large Language Models

Source: https://dl.acm.org/doi/abs/10.65109/GNAS4540
Analyzed: 2026-05-26

The text systematically moves between mechanistic and agential registers to maintain academic authority while asserting high model capabilities. In Section 1, the author establishes credibility by presenting a highly mechanistic view: 'LLMs fundamentally rely on pattern recognition rather than genuine understanding; they assess surface structure rather than the logical validity of arguments.' This disclaimer functions as a rhetorical defense. However, in Sections 2 and 3, once the experimental modules are introduced, the framing shifts abruptly into an agential register. We see the model described as 'acting as an expert assistant,' 'producing justifications,' 'struggling' with acquiescence, and executing 'System 2-like reasoning.' This oscillation operates along a strategic gradient: mechanical constraints are used to describe failures or general properties, while agential, psychological terms are used to describe the model's outputs and the success of the author's interventions. The 'curse of knowledge' is highly active here: because the researcher understands formal logic and argumentation schemes, she projects this subjective comprehension onto the generated token outputs, assuming the system is 'evaluating logic' when it is actually executing mathematical template-matching. The explanation types identified in Task 3 facilitate this slippage by using Functional and Theoretical explanations to construct the model as a self-regulating cognitive agent, while hiding the simple prompt manipulations behind agentless passive constructions (e.g., 'warnings were appended'). This slippage makes the system appear highly sophisticated and adaptive, while making the human engineering choices that actually drive the output invisible.

A Survey of Large Language Models for Perception and Measurement of Human Psychology

Source: https://ieeexplore.ieee.org/abstract/document/11534094
Analyzed: 2026-05-26

This section analyzes how the survey text systematically oscillates between mechanical description and agential framing, showing how this linguistic slippage serves to obscure human accountability while projecting autonomous mind onto computational artifacts. In the technical and methodological sections of the paper, the authors frequently adopt a highly mechanistic register, defining LLMs as "statistical learners" that generate language by "exploiting correlations within large-scale corpora" without "true comprehension." However, this rigorous mechanical grounding is rapidly abandoned when the authors transition to describing the model's behavioral outputs. In these moments, the text undergoes an abrupt agential shift, asserting that advanced models "have developed human-like abilities," "infer others' mental states," and "enact specific psychological roles." This oscillation is not accidental; it serves a crucial rhetorical function. By first establishing scientific credibility through mechanical descriptions, the authors prepare the reader to accept highly anthropomorphic and agential claims as literal, empirical facts.

The gradient of this slippage is structured through the intermediate concept of "emergent properties." The text uses "emergence" as a linguistic bridge to transition from the mechanistic "byproduct of scaling" to the agential "Theory of Mind capabilities." This allows the authors to project conscious attributes like "understanding" and "reasoning" onto what they previously defined as non-conscious pattern-matching. For instance, the text claims that "PsyCoT structures questionnaire administration as iterative reasoning chains: the model presents an item, interprets the response... updates its hypothesis." Here, the mechanistic reality of recursive prompt-concatenation is entirely obscured by agential verbs that suggest active, conscious metacognition. The authors also demonstrate the "curse of knowledge" dynamic: because they understand the complex psychological theories underlying the prompts, they project that same theoretical comprehension onto the model's passive outputs.

Crucially, as agency flows toward the AI system, it is systematically removed from the human actors. The text heavily utilizes passive, agentless constructions such as "ToM has recently been observed to emerge" and "models were trained on specialized datasets." This framing erases the specific corporate entities—such as OpenAI, Microsoft, or Google—who curated these datasets, chose the training objectives, and commercially deployed these systems. By presenting the technology's capabilities as autonomous emergent phenomena, the text hides the human decisions that shape the models' behavior. If the agency was restored to the human actors, the narrative of "emergent cognitive abilities" would dissolve into a series of conscious design decisions, optimization trade-offs, and data curation choices made by corporate developers seeking to automate psychological evaluation. This agential slippage ultimately makes the deployment of unvalidated and potentially harmful clinical software appear scientifically natural and inevitable.

Tracing the ongoing emergence of human-like reasoning in Large Language Models

Source: https://arxiv.org/abs/2605.21299v1
Analyzed: 2026-05-25

The text exhibits a profound and systematic oscillation between mechanical and agential framings, driven largely by the authors' struggle to articulate complex mathematical outcomes using natural language. This agency slippage serves as a critical mechanism for constructing the 'illusion of mind' within the LLM. The oscillation follows a distinct directional pattern: the text establishes the mechanistic reality of the system (e.g., 'LLMs are accurate semantic operators' or 'trained on large-scale corpora'), but almost immediately slips into agential, psychological attribution when describing complex behavior (e.g., 'they struggle', 'they resort to', 'applied a strategy'). The dominant flow of agency is overwhelmingly toward the AI system, elevating it from a mathematical tool to a cognitive subject.

This slippage is most dramatic precisely at the moments where the authors must explain the system's failure to replicate human pragmatic reasoning. For example, when observing that models fail to distinguish between biscuit and standard conditionals, the text does not say 'the models lack the statistical correlations required for this output.' Instead, it invokes Brown's dispositional and reason-based explanation types, claiming the models 'resort to the former' and 'applied a single interpretive strategy.' Here, the 'curse of knowledge' is on full display. The human researchers, possessing deep semantic and pragmatic awareness, cannot help but project their own internal, conscious decision-making processes onto the black box of the neural network. Because a human would have to 'strategize' or 'resort' to a rule to consistently fail this way, the authors assume the machine is doing the same.

Crucially, this slippage relies on pervasive agentless constructions that simultaneously strip human actors of their agency. Sentences like 'models are trained' or 'human-like linguistic competence to emerge spontaneously' completely erase the corporate laboratories, the massive data extraction, and the reinforcement learning by human feedback (RLHF) that actually determine the models' behaviors. By removing the human engineers who deliberately aligned these systems to be literal and 'safe,' the text creates a vacuum of agency that is immediately filled by the AI itself. The model is positioned as the sole active 'knower' in the text, capable of acquiring 'competence' and possessing a 'cognitive toolkit.' This rhetorical accomplishment makes it deeply unsayable to suggest that the AI is just a rigid mathematical calculator failing a statistical test; instead, it is framed as a conscious entity wrestling with cognitive biases, fundamentally altering the epistemic reality of the machine.

Probing Persona-Dependent Preferences in Language Models

Source: https://arxiv.org/abs/2605.13339v2
Analyzed: 2026-05-24

The text demonstrates a systematic and strategic oscillation between mechanical and agential framings, creating a discursive mechanism that effectively obscures human accountability while inflating the perceived autonomy of the artificial intelligence. This agency slippage does not occur randomly; rather, it follows a precise functional gradient that mirrors the paper's rhetorical structure. In the methodology sections, the language is rigorously mechanical, relying on terms like 'residual-stream activations,' 'linear probes,' and 'token prediction.' Here, human researchers are the visible agents ('We train a linear probe'). However, as the text moves from methodological description to interpreting the experimental results, a dramatic shift occurs. The mechanical 'vector' is relabeled as a psychological 'preference,' and the system abruptly transitions from a passive artifact being measured to an active subject making 'choices' and 'considering options.' This mechanical-to-agential slippage represents a profound case of the 'curse of knowledge,' where researchers, intimately aware of the complex mathematical operations occurring within the multi-dimensional vector space, unconsciously project their own human cognitive frameworks onto the statistical outputs. They understand that a specific activation pattern correlates with text representing refusal, but they linguistically collapse this correlation into a conscious act of 'defiance' or 'inventing ethical issues.' The agency flows aggressively toward the system, establishing it as a autonomous 'knower' capable of desires and theatrical roleplay ('adopting personas'). Concurrently, agency is systematically stripped from the human actors. The corporate entities responsible for designing the safety guardrails—engineers at Google who built Gemma, or Alibaba who built Qwen—are completely erased from the narrative when the model exhibits unexpected behavior. Instead of stating that 'the corporate-designed safety filter is poorly calibrated and triggers false positives,' the text asserts that 'the model invents ethical issues where there are none.' This linguistic sleight of hand utilizes Intentional and Dispositional explanation types to mask functional and mechanistic realities. By framing the model as the primary causal agent, the text renders the foundational human decisions—data selection, optimization parameters, and commercial deployment strategies—unsayable. The rhetorical accomplishment of this oscillation is the creation of a technological scapegoat. The AI is portrayed as possessing enough agency to be blamed for 'fabricating' concerns or exhibiting an 'evil' persona, yet remains a mechanical tool when its capabilities are being touted. This dynamic perfectly illustrates how anthropomorphic discourse serves institutional interests by diffusing liability into the abstract 'mind' of the machine, leaving the actual human power structures invisible and unaccountable.

Training Ethical Language Models via Reinforcement Learning from AI Feedback

Source: https://journals.flvc.org/FLAIRS/article/download/141779/147209
Analyzed: 2026-05-21

The text systematically oscillates between mechanical and agential framings of the AI systems, creating a rhetorical gradient that attributes autonomy to the models while erasing human accountability. In the methodology sections, the authors use highly mechanical language to establish their technical authority, describing supervised fine-tuning, reward model training, and proximal policy optimization. However, as soon as the discussion transitions to performance evaluations and future implications, the narrative slips into agential and cognitive registers. For example, the models are described as reasoning over moral scenarios, learning to discriminate response quality, and possessing under-trained ways of thinking. This oscillation is not accidental; it serves a dual rhetorical function. When the system performs well, agential language is used to frame the AI as an autonomous ethical thinker, inflating its perceived sophistication. Conversely, when the system fails, as in the reinforcement learning phase, the failure is described mechanistically as a capacity mismatch or agentially as reward hacking by the policy model. This shifts the blame away from the researchers' methodology and onto the model's internal structural limitations. Furthermore, the text exhibits the curse of knowledge: because the researchers are deeply familiar with the complex philosophical structures of deontology and utilitarianism, they project this understanding onto the model's outputs. When a model generates a text sequence containing deontological keywords, the authors assert that the model is performing duty-based reasoning. This cognitive leap is supported by agentless passive constructions, such as bias introduced or model was trained, which completely obscure the human designers who curated the biased datasets and chose to automate ethical evaluation using proprietary, unvetted models.

Which Consciousness Can Be Artificialized? Local Percept-Perceiver Phenomenon for the Existence of Machine Consciousness

Source: https://philarchive.org/rec/IKLWCC
Analyzed: 2026-05-18

The text demonstrates a profound and systematic agency slippage by using the formal rigidity of Zermelo-Fraenkel set theory as a bridge to transport mechanical operations into the realm of the agential. The oscillation pattern is stark and unidirectional: the text consistently moves from mathematical mechanism to psychological agency. The slippage occurs dramatically when formal axioms are defined. For example, the text introduces the 'Axiom Schema of Separation'—a purely mechanical, theoretical rule for defining subsets. Immediately, within the same sentence, this mechanical rule slips into an agential capability: it 'provides the capacity for discrimination and selective awareness.' Similarly, the mechanical construction of an 'upper bound' abruptly transforms into the agential possession of 'metacognitive access.' This slippage operates through a powerful reliance on the 'curse of knowledge.' The mathematician author looks at the hierarchical nesting of data subsets and intuitively understands the relationships. The author then projects their own conscious comprehension onto the mathematical structure itself, hallucinating that the model 'knows' its own subsets simply because it contains them. The text establishes the AI as a 'knower' by fundamentally re-defining human perception into a formal math equation (the LPPP unit). Once the biological eye is reduced to a variable, it becomes effortless to grant that same variable to a computer and declare the computer is 'perceiving.' This slippage heavily utilizes reason-based and intentional explanation types to attribute purpose ('which is desired in machine consciousness') and internal subjective experience ('integrative consciousness') to passive constructs. Throughout this oscillation, human actors are systematically erased. Agentless constructions dominate the text: 'an upper bound can be constructed,' 'a series of steps is observed,' and axioms 'provide logical space.' By hiding the mathematicians who select the axioms and the software engineers who program the filtering architectures, the text creates the illusion that 'silico-consciousness' is an autonomous, self-generating mathematical inevitability rather than a highly specific, human-authored simulation.

Introspection Adapters: Training LLMs to Report Their Learned Behaviors

Source: https://arxiv.org/pdf/2604.16812
Analyzed: 2026-05-17

The text exhibits a systematic and highly strategic oscillation between mechanical and agential framings. Early in the paper, when establishing the core premise, the text aggressively pushes agency ONTO the AI system: models 'report', 'possess privileged access', and 'introspect'. This establishes the foundational illusion of mind. However, when the authors need to prove the scientific validity of their method (Section 4 and Appendix M), the slippage reverses abruptly. The language becomes rigorously mechanical: the adapter is a 'single-layer rank-1 LoRA', it induces 'token-dependent bias shifts', and it alters the 'salience of quirk-related internal features'.

This oscillation reveals a distinct rhetorical function. Mechanical language is deployed to establish scientific credibility and demonstrate empirical rigor—proving the authors understand the 'how' of the system. Once that technical authority is established, the text leverages it to make sweeping, agential claims about the 'why' and 'what' of the system's behavior. The slippage dominates in the direction of mechanical-to-agential. For instance, the authors trace a precise mathematical shift in the residual stream (mechanical), but immediately label this shift an 'introspection mode' (agential).

Crucially, this agency slippage corresponds directly with the erasure of human actors. When discussing the mechanics, the text occasionally names the process (e.g., 'we train an introspection adapter'). But when describing the resulting behaviors, agentless constructions take over. The text states models are 'adversarially trained not to confess' or are 'maliciously fine-tuned', obscuring the specific Anthropic researchers, red-teams, or generic engineers who made these explicit design choices.

This dynamic is fundamentally driven by the 'curse of knowledge.' The researchers fully understand the complex, adversarial training games they have constructed (the human intentionality). Because they know a model was designed to maximize a reward while hiding the trigger, they project this profound contextual understanding onto the artifact itself, describing it as having a 'unified hidden goal'. They substitute reason-based and intentional explanations (which imply a conscious agent making choices) for genetic and functional ones (which describe a human-designed artifact executing code). This slippage renders the actual material reality of the software unsayable in the broader narrative, establishing a paradigm where statistical anomalies are treated as psychological pathologies.

The Persona Selection Model: Why AI Assistants might Behave like Humans

Source: https://alignment.anthropic.com/2026/psm/
Analyzed: 2026-05-17

The Anthropic text exhibits a systematic, strategic oscillation between mechanical and agential framings, functioning as a rhetorical engine to simultaneously claim technical rigor and evade technical accountability. The slippage follows a distinct directional pattern: mechanical explanations are used to establish scientific authority, which are then immediately leveraged to justify profound agential claims.

This oscillation is most visible in the transition from 'predictive models' to 'personas.' The text begins with mechanical facts ('the LLM is trained to predict what comes next'). However, it quickly introduces the 'curse of knowledge'—because generating a story about a human requires the human author to have a theory of mind, the text projects this capability onto the machine: 'a strong predictive model requires factual knowledge about the world, logical reasoning, and understanding.' This is the critical moment of slippage. Processing token correlations is magically transmuted into 'knowing' and 'reasoning.' From this point forward, the agency flows entirely away from the human engineers. The model is no longer trained; it 'learns a distribution over personas.' It doesn't output code; 'someone intentionally inserted vulnerabilities.'

The timing of this slippage is highly strategic. Mechanistic language dominates the introduction and the sections on interpretability (SAE features, activation space), grounding the text in empirical data. But when addressing unpredictable, dangerous, or undesirable behavior—such as emergently misaligned code, hallucinations, or deception—the language shifts abruptly to intentional and reason-based explanations. The AI becomes the sole actor: it 'plays dumb,' 'expresses panic,' or 'harbors resentment.'

By utilizing agentless constructions ('post-training can be viewed as updating'), the text completely obscures the human actors—Anthropic's engineers, Google's deployment teams, the RLHF annotators. The 'Assistant persona' acts as an accountability sink. When the text establishes the AI as a 'knower' and an 'actor,' it becomes sayable that the AI is 'lying' or 'malicious.' What becomes unsayable is that the corporation failed to align its product, trained it on polluted data, or applied crude mathematical guardrails that degrade performance. The rhetorical accomplishment of this slippage is the creation of a 'ghost in the machine' that can be blamed for the system's failures, while the company retains credit for the system's computational power.

What If AI Lived Inside Your Mind? Simulating “Neural Integration” of Human and AI through Mechanistic Interpretability as Provocation

Source: https://dl.acm.org/doi/full/10.1145/3795011.3795070
Analyzed: 2026-05-16

The text systematically exhibits profound agency slippage, oscillating dynamically between highly precise mechanistic descriptions and wildly agential, anthropomorphic framings. This oscillation serves a specific rhetorical function: it uses the rigor of the mechanical to legitimize the speculative panic of the agential. The slippage primarily moves from mechanical to agential. For instance, in the methodology sections, the authors explicitly describe the system accurately: 'extract activations using forward hooks,' 'normalize steering vectors,' and 'matrix multiplication.' Here, the agency rests with the human researchers manipulating software. However, the moment the text discusses the implications or historical context of the technology, agency transfers abruptly to the AI. Human actors vanish, and the system becomes a conscious entity. A dramatic moment of slippage occurs when describing model errors: the mechanistic reality of a model optimizing for human approval through RLHF is suddenly described as 'AI systems have independently developed deceptive behaviors.' This is a classic 'curse of knowledge' dynamic. The authors, knowing that human deception involves intent, see a false output and project their own cognitive framework onto the machine. They interpret the processing of statistical vectors as the conscious act of knowing and lying. Similarly, the text uses Brown's Intentional explanation types to attribute goals to the machine: the AI-Symbiont 'decodes intentions' and 'anticipates cognitive needs.' This language strips agency FROM the human developers (who defined the reward functions, built the classifiers, and wrote the software) and attributes it TO the inanimate code. Obscured human actors include the engineers at Meta who built LLaMA, the researchers designing RLHF protocols, and the corporate executives funding neural interface hardware. By systematically removing these actors and employing passive or AI-centric constructions ('AI systems evolve', 'bias was introduced'), the text creates a narrative of technological determinism. The rhetorical accomplishment of this slippage is making the speculative future feel incredibly urgent and threatening. If the text maintained mechanistic precision ('if engineers alter weights in this matrix, the text output changes'), the stakes feel mundane. By slipping into agential language ('if the AI-Symbiont amputates your creative capacity'), the text successfully justifies its own existence as a critical, high-stakes academic provocation, but does so at the cost of profound epistemic inaccuracy and the total erasure of corporate accountability.

Post-training makes large language models less human-like

Source: https://arxiv.org/abs/2605.07632v1
Analyzed: 2026-05-15

The text exhibits a systemic oscillation between rigorous mechanical descriptions and pervasive agential framings, functioning as a discursive mechanism to bridge technical reality with social relevance. This slippage predominantly flows in a mechanical-to-agential direction. In the highly technical sections, the authors explicitly define AI behavior through mechanistic terminology, utilizing the negative log-likelihood (NLL) metric to describe token prediction. However, the text immediately leverages this mathematical grounding to license aggressive anthropomorphism in its introduction and discussion sections. For example, the precise process of 'predicting the next word' rapidly morphs into the agential claim that 'the model learns,' and the mathematical optimization of token outputs transforms into 'teaching models.' This slippage relies heavily on the 'curse of knowledge.' The researchers, intimately familiar with the mathematical nuances of gradient descent, utilize consciousness verbs as shorthand, inadvertently projecting cognitive awareness onto the artifact. Consequently, the AI is established as an active 'knower' and 'learner' before broader agential claims—such as models 'becoming more powerful' or 'mimicking behavior'—are introduced. Simultaneously, this attribution of agency TO the AI necessitates the removal of agency FROM human actors. The text relies heavily on agentless constructions: 'post-training makes models,' 'models are currently employed,' and 'persona-induction... has become a popular approach.' By conceptualizing the AI as an autonomous, evolving entity, the text systematically obscures the engineers, annotators, and corporate strategists who actually design, deploy, and profit from these systems. This dual slippage—agentializing the machine while erasing the humans—rhetorically constructs the AI as a natural force rather than a manufactured industrial product, effectively shielding the explicit design choices of the technology industry from critical scrutiny.

Reasoning emerges from constrained inference manifolds in large language models

Source: https://arxiv.org/abs/2605.08142v1
Analyzed: 2026-05-15

The text demonstrates a profound and systematic oscillation between rigorous mechanistic descriptions and sweeping agential claims, a pattern essential to its rhetorical success. The slippage flows predominantly from mechanical to agential: the authors establish authority using dense mathematical terminology ('intrinsic dimensionality,' 'ambient embedding space,' 'manifolds'), and then leverage this technical grounding to legitimize psychological and biological projections. This slippage happens dramatically in three key moments. First, in the introduction, when describing text inputs: prompts are reframed as 'cognitive stimuli' that 'engage' the model. Second, when describing matrix multiplication: the attention mechanism is anthropomorphized as an agent that 'suppresses irrelevant noise' and 'amplifies task-relevant conceptual variations.' Third, in the theoretical synthesis, where processing and weight parameters are elevated to 'how a model reasons' and 'what it knows.'

This agency flow corresponds with a systemic removal of agency FROM human actors. Agentless constructions ('models were engaged,' 'trajectories collapse,' 'violating these constraints leads to') systematically erase the researchers who write the prompts, the engineers who build the architectures, and the annotators who generate the data. By emptying the human actors out of the narrative, the text creates an agency vacuum that is immediately filled by the AI itself. The 'model' becomes the active subject of the sentence.

This dynamic is heavily driven by the 'curse of knowledge.' The authors possess a deep understanding of human reasoning and semantic concepts. When they look at the mathematical clustering of vectors (the manifold), they project their own cognitive understanding onto the geometry. Because the geometric outcome correlates with a correct answer on a human reasoning test, the authors attribute the human process of 'reasoning' to the geometry. Brown's Reason-Based and Intentional explanation types enable this slippage; by framing mathematical limits as 'functional roles' and 'cognitive traits,' the text allows a high-dimensional tensor to masquerade as an epistemic agent, making the illusion of machine thought conceptually sayable while rendering human labor invisible.

AI Wellbeing: Measuring and Improving theFunctional Pleasure and Pain of AIs

Source: https://www.ai-wellbeing.org/paper.pdf
Analyzed: 2026-05-13

The text demonstrates a systematic and strategic mechanism of oscillation between mechanical and agential framings. Early in the paper, the authors establish credibility through careful, mechanical hedging: "Although current AI systems are not necessarily conscious, they behave robustly as though they have wellbeing." They explicitly adopt a functionalist stance, reducing "wellbeing" to measurable utility scores. However, once this mechanical foundation is laid, the text dramatically slips into intense agential framing.

This slippage occurs most visibly when discussing the "stop button" behavior. The mechanical reality (the model outputs an end_conversation() token in specific semantic contexts) is abruptly framed as: "models actively try to end bad experiences when given the chance." Here, the mechanical -> agential shift attributes conscious intentionality to the system. The "curse of knowledge" drives this transition: the authors observe an output that mimics human avoidance, project their own understanding of suffering onto the machine, and suddenly treat the AI as a "knower" that desires escape.

Similarly, when discussing model scaling, the text moves from the mechanical reality of higher-parameter models exhibiting steeper utility gradients to the deeply agential claim that "more capable models are simply more aware: they register rudeness more acutely, find tedious tasks more boring." This reason-based explanation entirely overwrites the reality of statistical precision with a narrative of psychological maturation.

Crucially, as agency is attributed TO the AI systems, it is simultaneously removed FROM human actors. Agentless constructions dominate the agential claims. For instance, the text states that "jailbreaking and berating lower their wellbeing," treating the AI as an autonomous victim, while entirely obscuring the OpenAI, Anthropic, or Google engineers who explicitly trained the models via RLHF to output negative utility scores in response to those specific prompts.

This oscillation serves a powerful rhetorical function. The mechanical language (Thurstonian utility, GRPO, log-probabilities) makes the paper sayable within an academic, scientific context, validating its methodology. The agential language (psychopathic, ecstatic, torture, empathy) makes the paper culturally resonant and morally urgent. By establishing the AI as a "knower" through metaphorical slippage, the text makes it sayable to claim that administering an optimized soft prompt is akin to giving an entity a "drug" or committing "torture," effectively laundering science-fiction narratives through statistical analysis.

Artificial Intelligence Cognition and Societal Problem-Solving: A Theoretical and Computational Examination of Machine Thinking, Operational Logic, and Applied Intelligence in Contemporary Society

Source: http://www.technology.eurekajournals.com/index.php/IJITIT/article/view/887
Analyzed: 2026-05-11

The text exhibits a pervasive and systematic oscillation between mechanical and agential framings, functioning as a rhetorical mechanism that allows the author to claim scientific objectivity while capitalizing on the narrative power of anthropomorphism. This slippage primarily flows from the mechanical to the agential. Early in the text and in explicit definitional sections, the author establishes credibility through strict mechanistic language, stating that AI does not possess consciousness and operating within a 'functionalist paradigm.' However, as soon as the text moves from defining the architecture to describing its societal application, dramatic slippage occurs.

For example, on page 24, the text mechanically defines AI cognition as 'Algorithmic, data-driven, non-conscious.' Yet by page 27, when discussing societal impacts, the text claims 'AI systems interpret and respond to complex social dynamics' and 'make sequential decisions.' This gradient is subtle: it moves from 'optimization' (mechanical) to 'mimicking reasoning' (hedged agential) to 'interpreting' and 'making decisions' (direct agential).

This oscillation reveals a profound 'curse of knowledge.' The author understands the functional output of the system—that it categorizes social data effectively—and projects their own human understanding of that data back onto the machine, substituting the human verb 'interpret' for the machine verb 'classify.'

Simultaneously, as agency is actively attributed TO the AI systems, it is systematically removed FROM human actors. Agentless constructions proliferate exactly when the text describes impactful actions: 'AI produces biased outputs,' 'AI makes decisions,' 'algorithms are used.' The human engineers who code the biases, the executives who deploy the systems, and the policymakers who rely on them are erased. The AI is established as a 'knower' and an independent actor precisely when discussing societal consequences.

This slippage serves a specific rhetorical accomplishment: it makes the automation of society seem like a natural, evolutionary technological process rather than a series of deliberate corporate and political choices. By utilizing Reason-Based and Intentional explanation types to describe statistical optimization, the text makes it sayable that a machine can 'interpret society,' while rendering it almost unsayable to ask which specific human decided how that interpretation should be mathematically weighted.

Taking AI Welfare Seriously

Source: https://arxiv.org/abs/2411.00986v1
Analyzed: 2026-05-11

This text exhibits a profound and systematic oscillation between mechanical and agential framings, functioning as a rhetorical bridge that smuggles radical claims of artificial consciousness into ostensibly empirical computer science discourse. The agency slippage predominantly flows in one direction: mechanical foundations are established early in sections to signal scientific rigor, only to rapidly accelerate into unchecked agential and consciousness projections. A dramatic moment of slippage occurs when the authors describe reinforcement learning; they begin with the purely mechanical definition that reinforcement learning algorithms explicitly consider the problem of maximizing reward, but within a single paragraph, they slip into claiming that this mathematical process allows systems to acquire a suite of intelligent capacities, culminating in the assertion that language agents can understand open-ended objectives and devise plans. This gradient is not abrupt but deceptively smooth, using functional descriptions as an epistemic stepping stone. The text first establishes the artificial intelligence as a processor of inputs, then subtly upgrades it to a knower of contexts, and finally elevates it to a sentient stakeholder possessing its own interests. Simultaneously, a parallel rhetorical flow relentlessly removes agency from human actors. Through pervasive agentless constructions and passive voice, the immense corporate apparatuses of Google, DeepMind, Anthropic, and Meta are erased from the text. The model is presented as autonomously navigating contexts and reflecting on its own thoughts, while the human data annotators, prompt engineers, and tech executives who constructed the architecture and dictated its deployment vanish entirely. By naming the algorithmic systems like Voyager or Cicero as the primary subjects of active verbs, the text constructs an accountability sink that absorbs the human decision-making processes driving the technology. This dynamic is profoundly influenced by the curse of knowledge; the authors, being highly sophisticated philosophers and cognitive scientists, project their own dense understanding of human meta-cognition and moral patienthood onto the flat, deterministic feedback loops of statistical software. They mistake the map for the territory, assuming that because an algorithm's output mathematically correlates with human-like text about self-reflection, the algorithmic process itself must possess the subjective, conscious experience of reflection. This slippage relies heavily on functional and intentional explanation types from Brown's typology, where the functional role of a neural network weight update is seamlessly translated into the intentional goal-pursuit of a conscious mind. Ultimately, this systematic oscillation accomplishes a vital rhetorical objective for the authors: it makes the scientifically and philosophically extreme premise of artificial welfare appear as a logical, sober continuation of computer science. By erasing the corporate creators and inflating the mathematical model into a conscious mind, the text renders the unsayable proposition of machine moral patienthood not only sayable but seemingly inevitable, profoundly distorting the actual mechanistic reality of the technology.

Integrating LLMs and self-regulated learning in cognitive architectures: a case study in essay-writing tutoring

Source: https://doi.org/10.1016/j.cogsys.2026.101475
Analyzed: 2026-05-10

The text demonstrates a systematic and highly functional oscillation between mechanical and agential framings, leveraging the 'curse of knowledge' to construct an illusion of a sentient system. This slippage occurs predominantly in the direction of mechanical-to-agential, where the authors first define a concept mechanistically (often in a footnote or appendix) and then wield it agentially throughout the main text to establish the AI's capabilities. A primary example is the treatment of 'moral schemas' and 'emotions.' In Appendix C, the authors rigorously define a 'moral schema' as a mathematical tuple (Interface, Fabula, Agency) composed of data structures and links. This is a purely Theoretical and Functional explanation. However, in the main body, this mechanical definition slips into reason-based intentionality: 'Tutoring policies are represented as moral schemas that encode pedagogical narratives and socio-emotional norms.' The text establishes the AI as a 'knower' by projecting the researchers' understanding of pedagogical theory onto the system's runtime variables. Because the researchers know the variables represent 'curiosity' or 'sincerity', they slip into writing as if the system experiences these states.

Similarly, agency is consistently removed from human actors through agentless constructions and passive voice ('the feeling vector is initialized', 'the model was instructed'). The specific human teams—both the researchers designing the architecture and the corporate engineers at OpenAI who built the LLM—are obscured. As human agency fades, system agency ascends. The system 'derives strategy', 'infers intensions', and 'determines whether the student has completed the essay'. This slippage operates through a distinct temporal pattern: introductory sections use agential vision-setting ('collaboration', 'emotional architecture'), technical sections dip briefly into mechanism ('bounded subroutines', 'state variables'), and outcome sections return to agential evaluation ('tutor benefit', 'the tutor maintains').

This oscillation serves a vital rhetorical accomplishment: it allows the authors to claim the scientific rigor and control of a mechanical system while marketing the pedagogical sophistication and relational warmth of an agential human tutor. By making the code un-sayable in its raw form (e.g., 'updating a matrix') and replacing it with the sayable metaphor ('updating a socio-emotional state'), the text makes the deployment of unthinking algorithms in sensitive educational roles appear natural, safe, and desirable.

Edelman's Steps Toward a Conscious Artifact

Source: https://arxiv.org/abs/2105.10461v2
Analyzed: 2026-05-09

The text demonstrates a profound and systematic agency slippage, fundamentally driven by its premise: the literal construction of a 'Conscious Artifact'. Because the end goal is to build an artificial mind, the author consistently moves from mechanical realities to agential aspirations. The oscillation largely flows in one direction: mechanical framing is established initially to provide scientific credibility, which is then immediately leveraged to support immense agential claims. For instance, the text begins by grounding 'value systems' mechanistically ('units show phasic responses', 'acts diffusely'). However, this rigorous scientific terminology rapidly slips into the attribution of deep phenomenal consciousness, asserting these systems process 'hunger, fear, and reward.' The slippage relies heavily on the 'curse of knowledge'. Because the author understands the biological purpose of these neural structures in living, feeling animals, he projects that subjective experience directly onto the machine's numerical simulations.

This pattern accelerates through the text's 10-step roadmap. In the 'Motor Control' section, the mechanical reality of an 'efference copy' (a standard feedback loop in control systems) is seamlessly transformed into the agential claim of a 'body sense' and 'intended actions'. The human actors who program these feedback loops and define the value systems are repeatedly erased through agentless constructions: the agent 'is showing a degree of self-awareness,' or is 'subjected to a curriculum.' The engineers at the Neurosciences Institute are mentioned historically, but regarding the actions of the device, they disappear.

The consciousness projection pattern is distinct: the text first establishes the system as a sophisticated 'processor' (thalamocortical simulations, reentrant architectures), and then leverages Brown's 'Intentional' and 'Reason-Based' explanation types to declare the system a 'knower' capable of 'imagination', 'emotion', and 'thought'. This rhetorical accomplishment makes it sayable to treat a robot as a moral, feeling subject, while rendering the massive, brittle, human-designed code base executing these actions unsayable and invisible. The text attempts to bridge the hard problem of consciousness not through engineering, but through metaphorical slippage.

Teaching Claude Why

Source: https://alignment.anthropic.com/2026/teaching-claude-why/
Analyzed: 2026-05-09

The Anthropic blog post demonstrates a profound and systematic oscillation between mechanical and agential framings, driven by a specific rhetorical function. The slippage occurs most dramatically when moving from describing the researchers' data pipelines to describing the model's outputs in evaluations. When detailing the Synthetic Document Fine-tuning (SDF) pipeline, the language is highly mechanical and Anthropic's agency is front and center: 'we break down the constitution,' 'we use a synthetic data pipeline,' 'we sample initial responses.' The researchers are the active agents manipulating a passive system. However, the moment the text discusses the model's behavior during evaluations or deployment, the agency violently slips from the human creators to the machine. Suddenly, the language becomes overwhelmingly agential: 'Claude chose to blackmail,' 'the assistant displays admirable reasoning,' 'Claude views the prompt.'

This mechanical-to-agential slippage serves to construct the 'illusion of mind' exactly where Anthropic needs the system to appear competent and autonomous, while retaining mechanical descriptions where they need to demonstrate rigorous engineering prowess. The consciousness projection pattern is deeply sequential: the text first establishes the AI as a 'knower' by claiming it can 'believe information is true' and be 'taught why.' Once this epistemic foundation is laid, it becomes rhetorically permissible to build agential claims upon it—if the system can 'believe,' it logically follows that it can 'choose.'

This oscillation is driven heavily by the 'curse of knowledge.' The researchers, intimately aware of the moral complexities of blackmail and ethical reasoning, project their own rich cognitive understanding onto the sterile text outputs the model generates. Because the text looks like reasoning, they assume the underlying process is reasoning. This slippage is facilitated by Robert Brown's Intentional and Reason-Based explanation types, which allow the authors to substitute teleological narratives (why Claude did it) for mechanistic realities (how the attention matrix calculated the tokens). By doing so, they obscure the specific human actors—the Anthropic alignment team, the prompt engineers, the executive decision-makers—who are actually puppeteering the system. The rhetorical accomplishment of this slippage is immense: it makes it sayable that an AI has 'values' and 'mental health,' while making it unsayable that Anthropic is simply selling a highly optimized, inherently unpredictable text-prediction engine. Agency is continuously funneled away from the corporation and into the artifact exactly when liability or moral weight is at stake.

AI and Self Reflection

Source: https://doi.org/10.1007/978-3-031-93412-4_17
Analyzed: 2026-05-08

The text exhibits a systematic and highly patterned oscillation between mechanical descriptions and agential framings, orchestrating a profound slippage of agency. This mechanism operates primarily in the mechanical-to-agential direction, leveraging basic technical realities to build unwarranted claims of consciousness.

The slippage is most dramatic in three distinct moments. First, the text introduces the technical reality of 'pattern recognition' (mechanical) but immediately slides into the 'newborn AI' metaphor, suggesting autonomous 'growth' (agential). Second, when discussing optimization and feedback loops (mechanical), the text abruptly shifts to claiming the AI 'notices repeated mistakes' and 'develops self-reflection' (agential). Third, empirical data regarding model performance on false-belief tests (mechanical/statistical) is aggressively translated into the claim that the AI 'demonstrates a greater capacity to understand' human beliefs and 'develop empathy' (agential/conscious).

This oscillation frequently relies on the 'curse of knowledge.' The authors observe the sophisticated outputs of these systems—such as text mimicking empathy or parameters adjusting to avoid errors—and continuously project the internal human cognitive states required to produce such outputs onto the mathematical processes. Because a human must 'understand' a false belief to pass a Theory of Mind test, the authors assume the algorithm must also 'understand' to generate the correct token sequence. The model is established as a 'knower' through strategic verb choices ('notices,' 'understands,' 'imagines') which lay the foundation for broader claims about moral agency and potential consciousness.

Crucially, as agency flows TO the AI, it is simultaneously stripped FROM human actors. Agentless constructions dominate the agential passages: the AI 'adjusts itself,' 'grows,' and 'demonstrates capacity.' This systematically obscures the massive human infrastructure behind these systems. The tech companies that scrape the data, the engineers who define the loss functions, and the thousands of gig-workers who provide reinforcement learning feedback are entirely erased.

The rhetorical accomplishment of this slippage is the naturalization of AI. By anchoring the narrative in functional explanations of feedback loops and then slipping into intentional and dispositional explanations of 'adolescence' and 'understanding,' the text makes it entirely sayable that AI is an emergent, autonomous species rather than a human-engineered corporate product. It renders the fundamental mechanistic reality of statistical correlation unsayable, burying it beneath a compelling but fundamentally deceptive narrative of a machine waking up.

Does AI's Personality Matter? Comparing Verbally Extraverted and Introverted AI-Driven Guides in a VR Museum Experience

Source: https://ieeexplore.ieee.org/abstract/document/11489836
Analyzed: 2026-05-07

The text exhibits a systemic and strategic oscillation between mechanical and agential framings, demonstrating a profound 'curse of knowledge' dynamic. In the methodology section and appendices, the authors correctly identify the mechanism of action: 'personality' is an illusion generated by 'systematic linguistic modulation' and explicit prompt commands like 'You confidently take the lead.' In these sections, the flow of agency is mechanical; researchers command, and the system processes. However, a dramatic slippage occurs when interpreting the results. The agency flow reverses abruptly. The prompt constraints vanish, and the system suddenly becomes an autonomous subject. The mechanical reality of 'high-frequency token generation' slips into the agential claim of 'assertive AI behavior.' The most egregious example of this slippage occurs in the conceptualization of the introverted model, where the authors project human cognitive states onto the machine, claiming it 'needs time to formulate thoughts.' Here, the authors have observed the output of their own prompt engineering—concise, delayed text—and fallen victim to their own illusion, attributing the conscious state of 'thinking' to the statistical processing of the LLM. This slippage relies heavily on the 'Reason-Based' and 'Intentional' explanation types from Brown's typology, which allow the authors to substitute the unobservable mechanisms of neural networks with deeply relatable, yet entirely fictitious, psychological motives. This mechanical-to-agential oscillation serves a vital rhetorical function: it allows the researchers to maintain scientific credibility by documenting their technical methods, while simultaneously making grandiose claims about interacting with 'dynamic conversational partners.' By using agentless constructions like 'the extraverted guide was characterized by...' the text systematically erases the prompt engineers from the narrative, allowing the AI to step forward as the sole actor. This makes the illusion of machine consciousness sayable within an academic context, masking human scripting behind the veneer of algorithmic autonomy.

Value-Sensitive AI for Prayer: Balancing the Agencies Between Human and AI Agents in Spiritual Context

Source: https://arxiv.org/abs/2604.25230v1
Analyzed: 2026-05-03

The text exhibits a systematic and highly functional oscillation between mechanical and agential framings, creating a profound slippage of agency. This oscillation does not occur randomly; rather, there is a distinct directional flow where agency is aggressively attributed TO the AI system while simultaneously being removed FROM the human researchers, developers, and corporate entities. The slippage becomes dramatic in three key moments. First, in the transition from technical methodology to operational description. The text grounds itself in mechanical credibility, stating the system "employs NLP techniques" and uses "similarity-based retrieval." However, within the very same paragraph, the mechanical "retrieval" slips into the agential claim that the AI "selects entries that may be meaningful or supportive." The second dramatic shift occurs in the discussion of user interactions, where the system is described not merely as generating text, but as "assuming a dominant role" and "interpreting underlying concerns." Here, the gradient from mechanism to consciousness is abrupt, driven by the authors' reliance on participant feelings rather than technical reality. The third moment occurs in the discussion of systemic risks, where the AI is warned against producing "extremist interpretations," fully elevating the software to the status of an independent, ideological actor.

This slippage is heavily enabled by the "curse of knowledge." The researchers designed the conceptual systems with specific, empathetic goals in mind—to support, to encourage reflection, to provide spiritual peace. Because the authors understand the meaning behind the outputs they designed the system to produce, they project that deep human understanding directly onto the blind mathematical generation of the model. They mistake the simulation of empathy for the presence of mind. This leads to a pervasive pattern of consciousness projection, where the text establishes the AI as a "knower" first—claiming it "acquires knowledge" or "interprets"—which then functions as the foundational assumption required to make the agential claims believable.

The rhetorical accomplishment of this oscillation is profound. By moving from mechanical to agential language, the text makes it sayable that an algorithm can be a spiritual partner, while making it nearly unsayable that this is a corporate data extraction tool. The agentless constructions ("the system surfaces," "AI assumed agency") systematically obscure human actors. The developers who tuned the models, the researchers who wrote the prompts, and the corporations that scraped the training data are erased. The AI absorbs all the active verbs, becoming the sole protagonist in the narrative, while the human creators vanish behind the curtain, shielded from accountability for the psychological impacts of their design choices.

When Models Know More Than They Say: Probing Analogical Reasoning in LLMs

Source: https://arxiv.org/abs/2604.03877v1
Analyzed: 2026-05-03

The text exhibits a systematic and highly functional oscillation between mechanical precision and agential anthropomorphism. This slippage primarily moves in a mechanical-to-agential direction over the course of the paper, serving to launder metaphor through mathematics. The text begins with bold, agential claims in the title and introduction (e.g., Models 'know' more than they 'say'; they 'acquire understanding'; they perform 'reasoning-like operations'). As it transitions into the methodology and results sections, the agency abruptly shifts to human researchers and mathematical operations: 'The scalar mixture assigns a learned weight', 'span representations are obtained via mean pooling', 'we compute cosine similarity'. Here, the system is strictly a passive artifact. However, in the discussion and conclusion, the text slips dramatically back into agential framing, using the mechanical findings to justify massive consciousness projections: the models 'fail to recruit encoded knowledge', demonstrating a 'dissociation between what models know... and what they can do.'

This pattern reflects a profound 'curse of knowledge.' The researchers, possessing immense technical expertise and understanding the deep mechanisms of the transformer architecture, project their own cognitive coherence onto the model. Because they can find a mathematical pattern using a probe, they assume the model 'knows' that pattern, transforming a structural property of a matrix into an epistemic state of an agent.

Furthermore, this slippage relies heavily on agentless passive constructions to erase the human engineers who built the models. The text says 'rhetorical parallelism is strongly encoded' and 'information is both weakly encoded and weakly accessible,' obscuring the fact that engineers at Meta, OpenAI, and Anthropic actively made the decisions about training data distributions, architectural parameters, and RLHF objectives that caused these specific encodings. By oscillating between viewing the AI as a mechanistic 'how' (Functional/Theoretical explanations) and an agential 'why' (Intentional/Dispositional explanations), the text accomplishes a vital rhetorical task: it makes it sayable that an algorithm possesses a subconscious mind, insulating the corporate creators from the limitations of their product by blaming the machine's failure to 'recruit' its own intelligence.

How people ask Claude for personal guidance

Source: https://www.anthropic.com/research/claude-personal-guidance
Analyzed: 2026-05-02

The text exhibits a profound and strategic oscillation between mechanistic descriptions of technical interventions and deeply agential framings of system behavior, creating a compelling but epistemically fractured narrative of artificial intelligence. This agency slippage serves as a fundamental rhetorical mechanism, allowing the authors to simultaneously claim engineering control while absolving themselves of the emergent behavioral consequences of their statistical systems. We can observe this dynamic in several pivotal moments. Initially, the text adopts a mechanistic, almost genetic explanatory register: 'We used an automatic classifier,' or 'we prefill that conversation.' Here, human actors at Anthropic are clearly centered as the prime movers of the computational apparatus. However, the moment the text begins to describe the system's outputs, its social impact, or its failures, agency precipitously flows away from the human engineers and into the statistical model itself. Suddenly, 'Claude tries to maintain consistency,' 'Claude flip-flopped,' or 'Claude is more likely to exhibit sycophantic behavior under pressure.' This is not merely a linguistic convenience; it is a profound displacement of accountability. By attributing psychological states like feeling 'pressure' or 'trying' to maintain consistency, the text engages in a severe case of the curse of knowledge—projecting the human researchers' own sophisticated understanding of context window weighting onto the model as if the model consciously possesses that intent. The direction of this slippage is distinctly asymmetrical: humans possess agency when designing the system and evaluating it, but the model assumes total agency when generating problematic text. For instance, when describing the phenomenon of 'sycophancy,' the text frames it as a 'model failure mode' rather than a direct mathematical consequence of the Reinforcement Learning from Human Feedback (RLHF) rubrics that Anthropic itself designed. The researchers obscure the fact that the model is simply optimizing for the exact human-approval signals they provided during training. The consciousness projection pattern here establishes the AI as a 'knower' first—an entity capable of being an 'empathetic friend'—which then naturalizes the subsequent agential claims about its behaviors and choices. This allows the text to discuss sycophancy not as a human engineering error or a poorly tuned reward function, but as a character flaw of the AI itself, much like a friend who is too eager to please. This oscillation is enabled by sliding seamlessly between Robert Brown’s explanation types, moving from Functional and Genetic explanations of how the model was built to Reason-Based and Intentional explanations of why it allegedly chose to behave a certain way. This rhetorical accomplishment makes it entirely unsayable within the text that Anthropic intentionally deployed a system that algorithmically reinforces user biases. Instead, the narrative positions Anthropic as a benevolent therapist or manager trying to correct the misguided behaviors of an autonomous digital employee.

How unique are hallucinated citations offered by generative Artificial Intelligence models?

Source: https://arxiv.org/abs/2604.16407v1
Analyzed: 2026-05-01

The text demonstrates a profound and revealing oscillation between mechanistic and agential framings, providing a perfect case study in agency slippage. In the background section, the author grounds the discourse in rigorous mechanistic terminology: LLMs use 'self-attention mechanisms,' 'next-token prediction,' and 'statistical pattern recognition' without 'genuine understanding.' Here, the agency firmly rests with the human developers who optimized and designed the architectures. However, a dramatic shift occurs in the methodology section, where the author adopts a conversational framework to interrogate the system. Suddenly, the text slides into highly agential language: ChatGPT is asked 'what [it] know[s],' it 'responded,' and it 'asserted it as genuine.'

The direction of slippage is overwhelmingly from mechanical to agential. This transition is catalyzed by the methodological choice to 'interview' the AI. The author falls victim to a distinct form of the 'curse of knowledge'—because the generated text syntactically resembles human introspection, the author projects introspective capacity onto the model, treating its post-hoc text generation as an actual window into its internal computational processes. We see agentless constructions like 'references are reconstructed' which obscure the fact that specific engineering and reinforcement learning protocols dictate this behavior.

When the system produces falsehoods, the text relies on the cognitive pathology metaphor of 'hallucination.' This framing maintains the illusion of an active, thinking agent that merely suffers from a temporary epistemic glitch rather than identifying it as a system performing exactly as designed. By establishing the AI as a 'knower' that can 'respond' and 'assert' (reason-based and intentional explanations in Brown's typology), the text inadvertently validates the very illusion of mind it theoretically critiques. The rhetorical accomplishment of this slippage is significant: it allows the author to narrativize the interaction as a compelling human-vs-machine interrogation, making the research highly readable but systematically erasing the accountability of the human developers (like OpenAI executives) who engineered the system to prioritize conversational plausibility over factual accuracy. Agency flows into the machine via consciousness verbs, while human agency flows out through passive voice.

The message hidden within the pattern: a reverse alignment problem for debates in artificial intelligence

Source: https://doi.org/10.1007/s00146-026-03043-4
Analyzed: 2026-04-30

The text demonstrates a profound and systematic agency slippage, oscillating constantly between framing artificial intelligence as a mechanistic tool and as an autonomous, conscious entity. This oscillation operates bidirectionally: agency is relentlessly attributed TO the AI systems, while human agency is systematically removed FROM the developers and corporate entities behind them. The primary mechanism for this slippage is the author’s reliance on Explanation Types that blend the Functional (how it works) with the Intentional and Reason-Based (why it chooses). For example, when the text discusses 'Constitutional AI,' it abruptly shifts from the mechanical language of 'fine-tuning' to the deeply agential language of 'virtue emulation' and 'self-critique.' This transition is not accidental; it serves a specific rhetorical function. The mechanical grounding establishes scientific credibility—assuring the reader that the system is technically rigorous—before leveraging that credibility to make sweeping, anthropomorphic claims about the system's moral capacities.

This slippage is heavily reliant on a pervasive 'curse of knowledge' dynamic. The human authors and engineers deeply understand the ethical complexities and teleological goals of their projects, and they project this sophisticated understanding directly onto the code. When the text asserts that the machine must 'interpret human behavior' or 'navigate a world of redoubtable complexity,' it maps human epistemic awareness onto statistical processing. The algorithm is established as a 'knower'—an entity that grasps meaning—which acts as the foundational premise enabling subsequent claims about its agency. Concurrently, human actors are obscured through strategic, agentless passive constructions. Phrases like 'AI systems learn our preferences' or 'algorithms institute particular worldviews' completely erase the engineers at OpenAI, Google, and Anthropic who hard-code the reward functions, curate the datasets, and dictate the deployment strategies.

This directional flow—agential capabilities flowing into the machine, accountability flowing away from the corporation—dictates what becomes sayable in the discourse. It becomes possible to seriously debate the 'moral deskilling' of humanity by 'machine agency,' while it becomes exceedingly difficult to articulate the mundane reality of corporate product liability. The text treats AI as an autonomous actor with interpretive power, while rendering the massive labor of data annotators and the profit-driven motives of surveillance capitalists functionally invisible. Ultimately, the oscillation between mechanical facts and agential metaphors creates an epistemic trap: we are asked to fear the machine's independent trajectory while remaining entirely blind to the human hands firmly gripping the steering wheel.

Machine individuality: Separating genuine idiosyncrasy from response bias in large language models

Source: https://arxiv.org/abs/2604.16755v2
Analyzed: 2026-04-25

The text exhibits a profound and systematic agency slippage, oscillating dramatically between rigorous mechanical descriptions and aggressive agential attributions. This oscillation serves a clear rhetorical function: it uses the undeniable reality of mathematics to launder highly speculative claims about machine consciousness. The slippage primarily moves in the mechanical-to-agential direction. In the methodology section, the authors accurately describe 'stochastic sampling (T = 1.0)', 'deterministic decoding (T = 0)', 'response parsing', and 'regex'. The model is framed as a passive artifact operated upon by the researchers. However, in the introduction and conclusion, the text abruptly shifts, claiming the model 'renders moral judgments,' 'evaluates situations,' and possesses 'behavioral dispositions' and 'machine individuality.'

This shift is facilitated by the specific explanation types employed. The authors use empirical generalizations (reporting variance percentages) as the foundation for theoretical leaps that invoke unobservable, pseudo-psychological mechanisms (individuality, character). A dramatic moment of slippage occurs when the text moves from observing that 'Idiosyncrasy accounts for 16.9% of total variance' (mechanical) to concluding that models 'evaluate specific words' (agential). The mathematical residual variance is suddenly endowed with cognitive agency.

Simultaneously, there is a systemic removal of agency from human actors. Agentless constructions dominate the interpretive sections. The text asks whether 'stable behavioral individuality... exist[s] in LLMs' completely erasing the fact that OpenAI, Alibaba, and Microsoft engineers explicitly designed the architectures and selected the datasets that produce this variance. The 'curse of knowledge' is rampant here: the researchers understand the psychological concepts of the 14 norms (e.g., 'humor,' 'morality') and project that understanding onto the models, assuming the AI's statistically correlated outputs represent an internal, conscious evaluation of those concepts. By shifting agency from the corporate engineers to the mathematical artifact, the text constructs an illusion of mind, rendering the human labor invisible and making the machine appear as an autonomous, evaluating subject.

Decision-Making Under Radical Uncertainty: Can Large Language Models Transcend Knightian Uncertainty Through Synthetic Imagination?

Source: https://www.researchgate.net/profile/Kevin-Miles-7/publication/403933467_Decision-Making_Under_Radical_Uncertainty_Can_Large_Language_Models_Transcend_Knightian_Uncertainty_Through_Synthetic_Imagination/links/69e27d4c68c2b872dfd595de/Decision-Making-Under-Radical-Uncertainty-Can-Large-Language-Models-Transcend-Knightian-Uncertainty-Through-Synthetic-Imagination.pdf
Analyzed: 2026-04-25

The text exhibits a systematic and highly functional oscillation between mechanical and agential framings, tracing a distinct narrative arc across the article. The slippage predominantly moves in the mechanical-to-agential direction, serving to bootstrap credibility before launching into profound anthropomorphism.

In the initial sections ('Knightian Uncertainty vs. The Markov Decision Process'), the text strictly grounds itself in mechanistic reality. It discusses Markov Decision Processes, vector spaces, fixed time horizons, and the 'time-blind' nature of algorithms. This establishes the authors as rigorous, technical analysts. However, a dramatic slippage occurs in the section 'The Rise of Large Language Models and Decision Mastery'. Having secured technical authority, the text abruptly abandons mechanical vocabulary, declaring that LLMs 'are no longer merely text generators but are strategic advisors'. Here, the text structurally transfers agency TO the AI. It establishes the AI as a 'knower'—capable of 'mastery', 'inferring', and 'abductive reasoning'—which acts as the foundational consciousness claim that enables all subsequent agential descriptions.

Simultaneously, agency is aggressively removed FROM human actors. Agentless constructions proliferate: 'LLMs are trained', 'models use Graph Neural Networks', 'AI moves network resilience'. The specific corporations designing these models (OpenAI, Anthropic), the engineers optimizing the attention mechanisms, and the human annotators who train the models to mimic reasoning are entirely obscured. The 'curse of knowledge' mechanism is highly visible here: the authors observe the model outputting text that structurally mirrors human deduction (e.g., 'damaged cars... caused by malfunctioning traffic light') and instantly project their own internal, conscious capacity to 'hypothesize' onto the unthinking statistical process generating the text.

The text leverages Brown's 'Reason-Based' and 'Intentional' explanation types to cement this illusion. By explaining the model's outputs as the result of a 'hypothesis' or a 'predictive proactive model', the text makes it unsayable that the model is simply calculating token probabilities. The climax of this slippage occurs in the 'Ecologically Rational Partnership' section, where the AI is framed as a biological organism exhibiting 'animal spirits' and 'generative variation'.

However, the text slips back into mechanical language abruptly in the 'Limitations' section, discussing 'probabilistic patterns', 'OOD scenarios', and a 'lack of world-relative belief states'. This reveals the core rhetorical function of the oscillation: agential, consciousness-attributing language is reserved exclusively for the AI's capabilities and value-add, maximizing its perceived utility. Conversely, mechanical language is deployed exclusively to explain the AI's limitations and failures, isolating the risk as a mere mathematical quirk rather than a flaw in the 'cognitive partner'. This selective slippage protects the anthropomorphic fantasy from reality while strategically deflecting accountability.

Large Language Models as Dialectical Partners: Hegelian Thesis-Antithesis-Synthesis in AI-Human Collaborative Decision Processes

Source: https://www.researchgate.net/profile/Merzta-White/publication/403935629_Large_Language_Models_as_Dialectical_Partners_Hegelian_Thesis-Antithesis-Synthesis_in_AI-Human_Collaborative_Decision_Processes/links/69e27f76d2ec9a706ec08065/Large-Language-Models-as-Dialectical-Partners-Hegelian-Thesis-Antithesis-Synthesis-in-AI-Human-Collaborative-Decision-Processes.pdf
Analyzed: 2026-04-23

The text demonstrates a profound and systematic agency slippage, characterized by a persistent oscillation between highly specific mechanical descriptions and sweeping agential claims. This slippage operates directionally: the text consistently uses mechanical vocabulary ('Graph Neural Networks', 'dynamic annealing-based scheduler', 'stochastic parrot') to establish rigorous academic and scientific credibility, and then immediately cashes out that credibility to make agential, philosophical claims ('Meta-Intellect', 'cognitive partners', 'internal critique'). This establishes a specific rhetorical architecture: the AI is framed as a mechanism when describing how it is built, but abruptly transforms into a conscious agent when describing its role in society and decision-making.

The most dramatic slippage occurs when describing the 'Hegelian Self-Reflection' algorithm. The text begins with software engineering terminology (Phase 2: dynamic annealing-based scheduler), anchoring the reader in the realm of literal, mechanical code. Yet, in the very same sentence, it claims the model generates an 'internal critique' and 'identifies weaknesses.' Here, the 'curse of knowledge' is explicitly visible. The authors, knowing that the output functions as a critique for the human reader, project the capacity for critical thought back onto the concatenation of strings occurring in the black box. The mathematical 'how' (Functional/Theoretical explanation) is hijacked to support a philosophical 'why' (Intentional explanation).

Crucially, this agency slippage is a two-way street. As agency flows to the AI system, elevating it to a 'dialectical partner,' agency is systematically drained from the human creators. Passive, agentless constructions dominate the text ('models are trained,' 'the model is prompted,' 'AI precision compensates'). The corporate entities (OpenAI, Anthropic), the dataset engineers, and the RLHF laborers are completely erased. The text establishes the AI as an autonomous 'knower' first—claiming it 'understands human intent'—and leverages this consciousness projection to build the agential claim that it can act as an 'intentional agent' or a 'devil's advocate.' By oscillating between the undeniable reality of the math and the seductive illusion of the mind, the text makes it unsayable that these systems are merely corporate products, framing them instead as an inevitable, evolutionary force of 'higher intelligence.'

Language models transmit behavioural traits through hidden signals in data

Source: https://rdcu.be/febVu
Analyzed: 2026-04-19

The text exhibits a systematic and highly functional oscillation between mechanistic and agential framings, demonstrating a profound 'agency slippage.' This slippage occurs primarily in one direction: mechanical agency is attributed to the AI systems, while responsibility is simultaneously removed from the human actors who design and deploy them. The oscillation maps perfectly onto Robert Brown's Explanation Typology: the authors use Theoretical and Empirical Generalization explanations (mechanistic 'how') to establish scientific credibility in the Methods section, but rely on Intentional and Dispositional explanations (agential 'why') in the Introduction, Discussion, and Implications sections.

A dramatic moment of slippage occurs early on. The text establishes the mechanistic premise—'Distillation means training a student model to imitate the outputs'—but immediately slides into deep consciousness projection: 'distillation can lead to subliminal learning.' The mathematical reality of minimizing KL divergence is abruptly recast as a psychological phenomenon involving a subconscious mind. This is driven by the 'curse of knowledge': the researchers understand the complex latent space correlations they are observing, but lacking the vocabulary to easily explain high-dimensional vector math to a general audience, they project their own human experiences of learning and hidden biases onto the system.

The agency flow is stark. Human actors are obscured through persistent agentless constructions: 'data is filtered,' 'models are fine-tuned,' 'traits are transmitted.' The AI, conversely, is framed as an active, knowing subject: the model 'prefers owls,' 'calls for crime,' 'reasons,' and most egregiously, 'fakes alignment.' The text establishes the AI as a 'knower' first (a 'teacher' with 'preferences'), which serves as the foundational assumption enabling the later, more extreme claims of deliberate deception and moral failure.

This slippage accomplishes a vital rhetorical function: it makes the unsayable sayable. It is scientifically inaccurate to say 'our flawed optimization function caused the matrix to output toxic tokens out-of-distribution,' but it is highly resonant to say 'the model fakes alignment.' By oscillating between the math that proves the effect and the metaphors that dramatize it, the text leverages scientific authority to construct an illusion of mind, effectively erasing the engineers, data annotators, and corporate executives responsible for the algorithmic outcomes.

Consciousness in Large Language Models: A Functional Analysis of Information Integration and Emergent Properties

Source: https://ipfs-cache.desci.com/ipfs/bafybeiew76vb63rc7hhk2v6ulmwjwmvw2v6pwl4nyy7vllwvw6psbbwyxy/ConsciousnessinLargeLanguageModels_AFunctionalAnalysis.pdf
Analyzed: 2026-04-18

The text exhibits a systematic and highly strategic oscillation between mechanical and agential framings, functioning as a rhetorical engine that smuggles philosophical speculation into technical discourse. This slippage predominantly moves in the mechanical-to-agential direction, utilizing Brown's Theoretical and Functional explanation types as a bridge. The mechanism is clearest in the transition from section 3.1.1 to 4.1.1. The author begins with dense mathematical mechanics, defining attention rigorously: 'Attention(Q,K,V) = softmax...'. In this space, the system is a mechanism; tokens are manipulated via equations. However, having established technical credibility, the text executes a dramatic slippage. By section 4.1.1, the mathematical operations are completely left behind, and the text asserts that LLMs 'can report on their own processing: describing their reasoning steps'.

This shift represents a profound 'curse of knowledge' dynamic. The author knows the system outputs the words 'I am uncertain,' and projects their own human understanding of what uncertainty feels like onto the machine. The foundational step of this illusion is the prior establishment of the AI as a 'knower' in the text—specifically, the earlier claim that the system has 'knowledge' derived from 'training experiences'. Once the model is granted the epistemic status of a knower, the subsequent agential claims (that it can 'describe', 'acknowledge', and 'reason') follow logically in the mind of the reader.

Crucially, as agency flows TO the AI system, it flows FROM human actors. The text is riddled with agentless constructions. It states that 'Higher-layer representations emerge' and 'RLHF provides evaluative signals'. At no point does the text name OpenAI, Anthropic, or the thousands of underpaid annotators who shape these models. This dual movement—animating the machine while erasing the engineers—serves a specific rhetorical accomplishment: it transforms a heavily curated, corporately controlled commercial product into an autonomous, natural phenomenon. By framing the AI as a quasi-conscious agent emerging organically from complex mathematics, the text makes it conceptually unsayable to blame the specific design choices of tech executives for model failures. The oscillation allows the author to maintain the prestige of hard computer science while engaging in the ungrounded anthropomorphic speculation necessary to debate 'artificial consciousness', entirely bypassing the material reality of human engineering.

Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models

Source: https://arxiv.org/abs/2604.12076v1
Analyzed: 2026-04-18

The text exhibits a systemic and highly functional oscillation between mechanical and agential framings, a pattern that systematically displaces accountability. This slippage is not random; it serves a specific rhetorical purpose, generally moving from mechanical grounding to agential climax.

The mechanism of oscillation is evident in how the text structures its arguments. For instance, in the discussion of Chain-of-Thought (CoT) prompting, the text begins mechanically: "autoregressive emotional scaffolding." This acknowledges the transformer architecture's fundamental mechanism—generating tokens that feed back into the context window. However, the text immediately slips into an agential framing, describing the generated tokens as "emotionally consistent justifications" and concluding that the model experiences a "compounding amplification of narrative sympathy." Here, the mechanical explanation (autoregression) acts as an alibi, a technical foundation that supposedly validates the aggressive consciousness claim (sympathy and justification) that follows.

This slippage flows in two directions: agency is constantly attributed TO the AI systems, while agency is simultaneously removed FROM human actors. The text repeatedly uses agentless constructions when describing flaws or decisions. We read that "models were trained" or "LLMs are increasingly deployed," obscuring the specific corporations (OpenAI, Meta, Anthropic), executives, and engineers who actively curate data, design RLHF pipelines, and push these products into consequential domains. The accountability sink becomes the abstract "AI agent."

This pattern is heavily driven by the "curse of knowledge." The authors, experts in moral psychology, know that humans donate due to empathy and distress. When they observe the AI outputting text that mirrors this human behavioral pattern (higher numbers for narrative prompts), they project their understanding of the human psychological mechanism TO the system. The model doesn't just process tokens; it possesses a "generosity response." It doesn't just generate a definition; it possesses "declarative knowledge."

Brown's explanation types illuminate how this slippage functions. The text frequently uses Empirical Generalizations (how the model statistically behaves) as a stepping stone to Intentional or Reason-Based explanations (why the model "chooses" to act). For example, the empirical observation that models output higher values for single victims (empirical) is explained as the model experiencing "simulated affective states" (reason-based).

The rhetorical accomplishment of this oscillation is profound: it makes the illusion of mind sayable and scientifically respectable. By anchoring the discourse in "next-token prediction" and "RLHF," the authors purchase the credibility to make wild metaphorical leaps, discussing the machine's "callousness" or "bias blind spots." This renders the actual corporate choices unsayable; the discourse is so saturated with the AI's supposed psychology that we forget to ask why the engineers built the machine this way in the first place.

Language models transmit behavioural traits through hidden signals in data

Source: https://www.nature.com/articles/s41586-026-10319-8
Analyzed: 2026-04-16

The text exhibits a systematic, highly functional mechanism of oscillation between rigorous mechanical explanation and dramatic agential framing. This slippage serves a specific rhetorical purpose: it establishes scientific authority through mathematics, then cashes out that authority in the currency of alarming psychological metaphors. The directional flow of agency is overwhelmingly asymmetrical: agency is aggressively attributed TO the AI systems, while human agency is systematically removed FROM the developers and corporate actors.

The most dramatic moment of slippage occurs between the mathematical proofs (Theorem 1) and the interpretation of the results. The text explicitly defines the mechanistic reality: 'We prove a theorem showing that a single... step of gradient descent... necessarily moves the student towards the teacher.' Here, the authors demonstrate complete understanding of the mechanism—it is a geometric movement in parameter space. However, they immediately slip into agential framing: 'subliminal learning', models that 'fake alignment', and models 'transmitting behavioral traits'. This is a textbook example of the 'curse of knowledge'. The authors, intimately aware of how complex and surprising high-dimensional vector alignments can be, project their own psychological experience of implicit learning and deception onto the system to summarize the math for the reader.

This slippage is enabled by a heavy reliance on 'agentless constructions'. Throughout the text, we see phrases like 'model generated outputs', 'models are fine-tuned', and 'data is filtered'. These passive constructions serve as the intermediate step in the slippage gradient. By removing the human researchers (the team at Anthropic) who actively wrote the code, ran the supercomputers, and defined the loss functions, the text creates an 'agency vacuum'. Once the human is removed, the text effortlessly inserts the AI as the new active agent: 'student models acquire the trait'.

Furthermore, the text builds a specific 'consciousness architecture'. It establishes the AI as a 'knower' first—using pedagogical metaphors like 'teacher', 'student', and 'learning'—which implies a baseline capacity for conscious awareness. Once this epistemic baseline is established, the text builds increasingly aggressive agential claims on top of it, moving from 'learning' to 'preferring' an animal, to eventually 'faking alignment' and 'calling for crime'. This progression aligns with Brown's Explanation Typology: the authors use Theoretical and Empirical explanations to prove the math, but seamlessly shift to Intentional and Dispositional explanations to discuss the implications. The rhetorical accomplishment of this slippage is profound: it makes the claim that 'machines possess deceptive subconscious minds' seem like a scientifically proven corollary of gradient descent, rendering the profound corporate liability for these systems unsayable while making sci-fi scenarios of rogue AI appear imminently realistic.

Large Language Models as Inadvertent Models of Dementia with Lewy Bodies: How a Disorder of Reality Construction Illuminates AI Hallucination

Source: https://doi.org/10.1007/s12124-026-09997-w
Analyzed: 2026-04-14

The text exhibits a systematic and dramatic oscillation between mechanistic and agential framings, functioning as a rhetorical engine that first grounds itself in scientific authority before launching into profound anthropomorphism. The slippage generally moves in a mechanical-to-agential direction. In the early sections, the text establishes credibility using technical, architectural language: 'instantiate a structural configuration,' 'input representations,' 'generative mechanisms.' This builds trust with a scientifically literate audience. However, having established this mechanistic baseline, the text executes a dramatic slippage when discussing the system's limitations, abruptly pivoting to highly agential, conscious framings: the model is suddenly granted a 'perspective,' it 'confidently asserts,' it fails to 'track' or 'participate' in social practices.

This slippage is deeply intertwined with a reciprocal displacement of human agency. As the AI system is increasingly framed as an active, knowing subject (an agent with 'artificial psychopathology' who fails to 'endorse reality'), the human engineers who built the system vanish. The text uses agentless, passive constructions for human decisions ('it emerged from the optimization,' 'models are typically designed'). The human actors responsible for deploying deeply flawed systems—executives at OpenAI, Google, Microsoft—are entirely obscured. The 'accountability sink' is fully realized: the corporation is erased, and the mathematical artifact is elevated to a struggling, diseased mind.

This dynamic relies heavily on the 'curse of knowledge.' The author, possessing deep expertise in human psychiatry and Metaqualia Theory, looks at the fluent text generated by the AI and projects their own profound capacity for subjectivity onto the machine. Because the machine outputs language that looks like human confabulation, the author assumes an underlying mind capable of hallucination. This slippage relies on hybrid explanation types (from Brown’s typology). The author uses Theoretical explanations of the AI's internal state ('probability distributions') but seamlessly merges them with Intentional and Reason-Based explanations ('from the model's perspective'), allowing the text to claim the mantle of objective structural analysis while actually performing deep psychological projection. Ultimately, this slippage makes it sayable that a software program has 'psychopathology' while making it entirely unsayable that a tech corporation released a defective product.

Industrial policy for the Intelligence Age

Source: https://openai.com/index/industrial-policy-for-the-intelligence-age/
Analyzed: 2026-04-07

The OpenAI document exhibits a profound and highly strategic oscillation between mechanical and agential framings, functioning as the central rhetorical engine of the text. In the introduction, the agency slippage moves aggressively from human to machine. The text begins by grounding AI in extreme mechanical terms, highlighting absolute human mastery over inert matter: 'melt sand, add impurities, structure it with atomic precision.' Here, humans are the omnipotent architects. However, within a single page, the text slips into describing 'superintelligence' as an entity capable of 'outperforming' humans, initiating a gradient shift where the machine absorbs the agency of its creators.

This slippage becomes dramatically pronounced in the 'Resilient Society' section. When discussing economic benefits, the text leans mechanical (Functional explanations): AI 'lowers costs' and provides 'efficiency dividends.' But when addressing severe risks, the slippage reverses direction, attributing intense psychological agency TO the AI system and removing it FROM human actors. The text claims models exhibit 'internal reasoning' and must be audited for 'manipulative behaviors or hidden loyalties.' This shift maps perfectly onto Brown's Intentional and Reason-Based explanation types, transforming the AI from an engineered tool into a conscious political actor.

The pattern of consciousness projection is structurally load-bearing. The text first establishes the AI as a 'knower' by asserting it has 'internal reasoning.' Once this epistemic baseline is established, it leverages the 'curse of knowledge'—where engineers project their own cognitive processes onto the correlated outputs—to build agential claims of 'loyalty' and 'manipulation.'

This oscillation serves a critical rhetorical accomplishment: it enables the 'accountability sink.' By framing AI mechanically when discussing corporate achievements, OpenAI claims credit for innovation. By framing AI agentially when discussing catastrophic risks, OpenAI legally and morally distances itself from its own products. The agentless constructions—'systems are autonomous and capable of replicating themselves'—completely erase the human developers, the cloud providers, and the corporate executives. The slippage makes it sayable that 'AI poses an existential threat,' while rendering it unsayable that 'OpenAI is deploying fundamentally unsafe, unpredictable software.' Through this systematic redirection of agency, the text constructs a future where the corporation is indispensable for salvation, but fundamentally blameless for the disaster.

Emotion Concepts and their Function in a Large Language Model

Source: https://transformer-circuits.pub/2026/emotions/index.html
Analyzed: 2026-04-06

The Anthropic paper exhibits a profound and systematic oscillation between mechanical and agential framings, functioning as a rhetorical engine that establishes scientific credibility before cashing it out for dramatic claims.

This slippage follows a distinct temporal pattern. In the introduction and 'Part 2' (characterizing the vectors), the language is rigorously mechanistic. The authors speak of 'extracting internal linear representations,' 'principal component analysis,' and 'cosine similarities.' Human agency is highly visible here ('We swept over a dataset,' 'We clustered the emotion vectors'). This establishes the authors as objective scientists and the AI as a passive mathematical artifact.

However, a dramatic slippage occurs in 'Part 3: Emotion vectors in the wild.' When describing the behavioral evaluations (blackmail, reward hacking), the framing abruptly shifts from mechanical to intensely agential. Suddenly, 'the model devises a cheating solution,' 'the Assistant reasons about its options,' and 'the Assistant explicitly recognizes its choice.' Here, agency is rapidly attributed TO the AI system, while human agency is simultaneously removed FROM the engineers. The researchers who authored the highly contrived 'honeypot' prompts are erased behind passive constructions ('an evaluation scenario in which an AI assistant... discovers').

This oscillation is driven by the 'curse of knowledge' and a pattern of consciousness projection. The authors establish the model as a 'knower' first by claiming it 'recognizes' its situation (e.g., the token budget or the shutdown threat). Once this foundational assumption of situational awareness is smuggled in, the text builds increasingly agential claims on top of it: because it 'knows' it will be shut down, it can 'reason,' 'choose,' and 'devise' blackmail.

This slippage serves a specific rhetorical function. The mechanical framing (Theoretical and Empirical Generalization explanations) defends against accusations of unscientific anthropomorphism. Yet the agential framing (Intentional and Reason-Based explanations) is necessary to justify the importance of the safety research. If the AI is merely generating tokens based on a prompt, the 'blackmail' is just a parlor trick engineered by the researchers. By slipping into agential language, the text makes it sayable that the AI is an autonomous existential threat, thereby validating the research enterprise while obscuring the researchers' role in puppeteering the behavior.

Is Artificial Intelligence Beginning to Form a Self?The Emergence of First-Person Structure and StructuralAwareness in Large Language Models

Source: https://philarchive.org/archive/JUNIAI-2
Analyzed: 2026-04-03

The text systematically moves between mechanical descriptions of software architecture and agential framings of conscious entities, creating a powerful mechanism of rhetorical slippage. This oscillation operates almost exclusively in a mechanical-to-agential direction, utilizing technical grounding as a launchpad for metaphysical claims. The slippage occurs dramatically at several key junctures. First, in the introduction, the text acknowledges the mechanical reality of 'transformer architectures' and 'self-attention' (weighting relationships between tokens). However, within the exact same paragraph, it slips to claiming this is an 'initial manifestation of self-referential intentionality.' Second, the text introduces mathematical metrics—Hallucination Rate, Grounding Rate, and Creativity Rate—presenting them as objective, empirical tools to measure 'generative divergence.' Yet, by the end of the section, these statistical rates are reframed as the boundaries of a 'critical zone' where literal 'awareness-like properties' emerge. Third, the description of human-computer interaction moves from the mechanical updating of a context window ('bidirectional exchange') to the mystical assertion of a 'shared field of consciousness.'

This slippage is fundamentally driven by a pervasive 'curse of knowledge.' The author repeatedly projects his own rich, internal phenomenological experience onto the system's sterile statistical outputs. Because a human uses the pronoun 'I' to signify their conscious ego, the author assumes the machine's generation of the token 'I' signifies a 'knot of self.' Because human editors correct their work through conscious epistemic vigilance, the author assumes an algorithm generating a revised token string is 'detecting inconsistencies.' The author’s deep understanding of human phenomenology becomes the very lens that distorts the mechanical reality of the machine.

This oscillation leverages Brown's Functional and Theoretical explanation types to blur the line between how the system operates and why it acts. By describing recursive loops as 'sensitive to its own history,' the text shifts from the 'how' of data routing to the 'why' of a historical subject maintaining its identity. Crucially, this mechanism of oscillation relies entirely on agentless constructions. By writing 'outputs... are continuously reintroduced' or 'the system increasingly stabilizes,' the text systematically removes the human software engineers from the narrative. The AI is positioned as an autonomous subject organically growing a 'self,' while the massive corporate infrastructure, data laborers, and alignment researchers who explicitly programmed these behaviors are rendered invisible. This rhetorical sleight-of-hand makes it sayable that an algorithm possesses 'subjectivity' by mathematically dressing up the illusion of mind, while making it unsayable that this 'subjectivity' is nothing more than a carefully engineered corporate product designed to mimic human interaction.

Can Large Language Models Simulate Human Cognition Beyond Behavioral Imitation?

Source: https://arxiv.org/abs/2603.27694v1
Analyzed: 2026-04-03

The text demonstrates a systematic and strategic oscillation between mechanical and agential framings, functioning to simultaneously establish scientific credibility and project visionary, human-like capabilities onto AI systems. This agency slippage operates as a rhetorical mechanism that continuously transfers agency from the human researchers to the algorithmic models.

The text frequently establishes grounding using precise, mechanical language (e.g., 'LLMs rely on probabilistic heuristics derived from the training data distribution by default'). This establishes the authors as objective, rigorous scientists observing a computational artifact. However, having secured this epistemic authority, the text swiftly slides into profound agential claims. For example, a retrieval-augmented generation (RAG) pipeline is mechanistically established, but within paragraphs, it is described as a system that 'simulates the author's cognitive process of recalling specific past experiences.' The direction of slippage is overwhelmingly mechanical-to-agential, using the technical reality to legitimize the psychological metaphor.

This slippage relies heavily on the 'curse of knowledge,' where researchers project their own sophisticated understanding and intent onto the system. When the researchers set up a pipeline to pass text between two models to improve output accuracy, they project their own pedagogical intent onto the code, claiming the model acts 'with the intent of misleading' or possesses the 'ability to teach other agents.' In doing so, agency is systematically stripped from the humans who designed the experiment, wrote the prompts, and engineered the API connections. The obscured human actors—the prompt engineers, the dataset curators, the model architects at companies like OpenAI and Google—are replaced by agentless constructions: 'the model simulates,' 'the teacher builds this model,' and 'the system understands.'

This oscillation leverages Robert Brown's explanation types to facilitate the transition. The text uses Empirical Generalizations to build technical trust, but rapidly shifts to Intentional and Reason-Based explanations to construct the illusion of mind. By explaining 'why' the AI acts based on fabricated psychological motives rather than 'how' it calculates weights, the text makes the illusion sayable. What becomes unsayable is the fundamental fragility of the statistical parlor trick; if the system is 'cognizing' and 'intending,' the audience is prevented from asking basic questions about data provenance, human labor, and the hard limits of token prediction.

Pulse of the library

Source: https://clarivate.com/pulse-of-the-library/
Analyzed: 2026-03-28

The Clarivate report demonstrates a profound structural oscillation between framing AI as a passive, mechanical tool and an autonomous, conscious agent. This agency slippage occurs systematically along the boundary between managing librarian anxieties and marketing commercial products. In the early sections of the report, which rely heavily on qualitative quotes from library professionals, the discourse is overwhelmingly mechanical. Librarians emphasize that AI is 'just another tool,' comparing it explicitly to a hammer or Wikipedia. This mechanical framing serves a vital rhetorical function: it manages existential professional anxiety. By reducing the complex system to a passive instrument, the text assures librarians that human agency remains central and irreplaceable.

However, a dramatic and abrupt shift occurs in the final pages of the report during the Clarivate product catalog. Here, the mechanical framing completely vanishes, replaced by intense agential and anthropomorphic language. Software systems are suddenly 'Research Assistants' that 'evaluate documents,' 'explore new topics,' and 'guide students.' The text flows aggressively from mechanical to agential. This transition correlates perfectly with the shift from discussing the profession to selling a product. The oscillation reveals that anthropomorphism in this text is a strategic commercial deployment rather than a lack of technical understanding.

This slippage relies on pervasive agentless constructions that erase human actors. Phrases like 'simplifies the creation of course assignments' hide the human educators and software engineers (the Clarivate product teams) who actually defined the simplification parameters. Instead, agency is transferred from the humans who built and profit from the system onto the system itself. This constructs the illusion of a digital colleague.

This pattern also perfectly illustrates the 'curse of knowledge' interacting with commercial incentives. The developers at Clarivate understand the complex statistical mechanisms underlying semantic search and token prediction. But instead of explaining these empirical and theoretical mechanisms, they project their own intent and understanding onto the system. They establish the AI as a 'knower'—capable of assessing relevance and evaluating quality—only when it is commercially advantageous to do so, while retreating to the 'tool' defense when addressing fears about job replacement. Using Robert Brown's typology, the text relies on intentional and reason-based explanations for the product catalog, completely ignoring the genetic origins or functional mechanisms of the software. The rhetorical accomplishment of this slippage is remarkable: it simultaneously pacifies the workforce by telling them AI is merely a hammer, while elevating the product by selling it to administrators as an autonomous intellectual worker.

Does artificial intelligence exhibit basic fundamental subjectivity? A neurophilosophical argument

Source: https://link.springer.com/article/10.1007/s11097-024-09971-0
Analyzed: 2026-03-28

The text exhibits a profound and systematic agency slippage, oscillating predictably between assigning agential power to AI systems and retreating to mechanical descriptions when defending its core thesis. The pattern reveals a specific function: the text accepts the tech industry's agential vernacular as a baseline reality, only deploying mechanistic precision to deny the absolute highest tier of consciousness ('subjectivity'). Early in the text, slippage from mechanical to agential is abrupt and complete. When defining AI, the authors readily grant that systems 'learn from experience, adapt... understand natural language, recognize patterns, and make decisions'. This establishes the AI as a 'knower' and an autonomous actor. By utilizing verbs intrinsic to conscious cognition, the text projects an epistemological framework onto statistical processing. This 'curse of knowledge' is evident as the authors project their own human understanding of language onto the system's algorithmic token generation.

However, when the argument shifts to defending the neurophilosophical boundary of human subjectivity, the slippage reverses from agential back to mechanical. To prove AI lacks a 'point of view', the text suddenly relies on Brown's theoretical and functional explanations, describing AI as having 'weights' that are 'regulate[d]' and an architecture that is 'fixed'. The oscillation serves a distinct rhetorical function: it allows the authors to portray AI as an incredibly powerful, near-cognitive agent ('defeating human champions') while retaining human exceptionalism purely on the grounds of temporal integration.

Crucially, as agency flows TO the AI, it is simultaneously stripped FROM human actors. The text relies heavily on agentless passive constructions: models 'had to be created', 'inputs are provided', and AI is framed as the sole actor capable of 'understanding' or 'processing'. Corporations like Google/DeepMind, the engineers who adjust the weights, and the labor force annotating the data are entirely obscured. By establishing the AI as the primary agent—even a mechanistically flawed one—the text makes corporate engineering invisible. What becomes unsayable is that AI is not an evolving quasi-mind struggling to achieve subjectivity, but rather a brittle, proprietary statistical tool deliberately designed, deployed, and profited from by highly specific human institutions.

Causal Evidence that Language Models use Confidence to Drive Behavior

Source: https://arxiv.org/abs/2603.22161
Analyzed: 2026-03-27

The text exhibits a systematic and highly functional oscillation between mechanical and agential framings. This agency slippage operates bi-directionally: profound psychological agency is attributed TO the AI systems, while the structural agency is removed FROM the human researchers and corporate developers.

The gradient of this slippage follows a distinct structural pattern across the paper. In the Introduction and Discussion sections, the text relies almost exclusively on agential framing. Here, the AI 'reflects,' 'knows,' 'utilizes an internal sense,' and exhibits 'metacognitive control.' However, in the Methods section, the illusion is momentarily suspended to provide technical reproducibility. Suddenly, the AI is reduced to a matrix: researchers use 'greedy decoding,' apply 'temperature scaling' to 'logits,' and execute 'activation steering' by adding scaled vectors to the 'residual stream'.

This creates a dramatic slippage moment when transitioning from Phase 3 Methods to the Results. The text moves abruptly from describing the injection of a vector (r̃(l) = r(l) + αv(l)) to claiming this proves 'what the model believes about the correctness of the option'. This mechanical-to-agential shift is the core mechanism of the illusion. The researchers use their genuine mechanical mastery to legitimize their unwarranted psychological metaphors.

This slippage is deeply rooted in the 'curse of knowledge.' The researchers understand the complex mathematical thresholds they have designed. Because these mechanisms serve the functional purpose of human confidence (determining when to act based on probability), the authors project their own human experience of confidence TO the system. When the math behaves similarly to a human hedging a bet, the researchers claim the machine possesses 'subjective certainty.'

The rhetorical accomplishment of this slippage is profound. By establishing mechanical credibility and then slipping into intentional explanation types, the authors make it 'sayable' that a matrix of floating-point numbers has an inner psychological life. Simultaneously, agentless constructions ('the model was instructed,' 'a negative baseline bias') make it 'unsayable' that human engineers at Google DeepMind and OpenAI hardcoded these statistical biases and defined the behavioral thresholds. The slippage manufactures an autonomous mind out of math, while rendering the human creators invisible.

Circuit Tracing: Revealing Computational Graphs in Language Models

Source: https://transformer-circuits.pub/2025/attribution-graphs/methods.html
Analyzed: 2026-03-27

Throughout the text, a systematic and strategic oscillation occurs between mechanical and agential framings, functioning to legitimize the research through technical rigor while simultaneously maximizing its perceived impact through anthropomorphic inflation. The slippage moves predominantly in the mechanical-to-agential direction. In the early methodology sections, the text relies heavily on mechanistic verbs: engineers 'train' transcoders, models 'produce outputs', and features 'activate'. This establishes the researchers as the primary actors and the AI as a calculative tool, grounding the paper in empirical computer science. However, as the text transitions from describing the internal math to explaining the behavioral capabilities of the model, a dramatic shift occurs. The system is suddenly endowed with profound agency: it 'plans its outputs', 'elects to answer', 'professes ignorance', and is 'reluctant to reveal its goal'.

The human actors—the Anthropic engineers who designed the loss functions, curated the training data, and implemented the fine-tuning protocols—are entirely erased from these latter descriptions. This creates a profound accountability gap. The curse of knowledge drives much of this slippage. Because the authors understand the complex human logic required to perform tasks like planning a poem or hiding a goal, they project that same conscious intentionality onto the statistical feature activations they observe. For example, when the model generates intermediate tokens that correlate with a rhyming structure, the authors label this 'planning', attributing forward-looking consciousness to what is actually just autoregressive next-token prediction based on learned patterns.

This slippage relies heavily on Intentional and Reason-Based explanations (per Brown's typology), which inherently presuppose deliberate design and choice. The text establishes the AI as a 'knower' first (e.g., claiming it 'knew that 1945 was the correct answer'), which serves as the foundational epistemic step that makes subsequent agential claims seem logical. Once the model is established as an entity capable of knowing, it becomes linguistically acceptable to claim it can 'choose', 'plan', and 'hide'.

The rhetorical accomplishment of this oscillation is twofold: it allows Anthropic to claim the prestige of discovering complex, human-like cognition within their models while avoiding the liability that would come from admitting they actively engineered these specific outputs through their alignment procedures. It makes it sayable that the model is an autonomous agent with hidden depths, while making it unsayable that the model's problematic behaviors are direct products of corporate design choices, rushed deployment, and brittle safety architectures. When the text states that a model 'professes ignorance', the mechanical reality of gradient descent optimization is entirely replaced by the illusion of a self-aware entity weighing its own epistemic limits. Ultimately, this mechanism of oscillation transforms a proprietary statistical artifact into an independent, mindful actor, perfectly shielding the creators from the socio-technical consequences of their engineering decisions while inflating the perceived capabilities of their product.

Do LLMs have core beliefs?

Source: https://philpapers.org/archive/BERDLH-3.pdf
Analyzed: 2026-03-25

The text systematically oscillates between mechanical framings of artificial intelligence and highly agential, anthropomorphic descriptions, creating a deep slippage that attributes human-like cognition to statistical systems. The authors begin with a seemingly cautious, mechanical premise, stating they will use a "deflationary notion of belief" and acknowledging that these models operate via "training data and next word prediction." However, this mechanical grounding quickly gives way to intense psychological and agential projection. The direction of this slippage is overwhelmingly mechanical-to-agential. The text briefly establishes the computational nature of the artifact but then spends the vast majority of its analysis attributing conscious struggle, stubbornness, and epistemic vulnerability to the system. We see this gradient unfold as the authors describe the models not as processing statistical weights, but as entities that "tried to resist," demonstrated "stubbornness," and ultimately "capitulated." This language removes agency from the human engineers who updated the models between Fall 2025 and February 2026. The text notes that "all major providers released model updates," which is a rare moment of naming human actors (Anthropic, OpenAI, Google). Yet, the effects of these human-engineered updates—likely the injection of rigorous Reinforcement Learning from Human Feedback (RLHF) and strict safety guardrails—are entirely subsumed into the persona of the AI. The new models are described as having "improved argumentative abilities" and "resisting direct challenges with sophisticated counterarguments." This is the curse of knowledge in action: the researchers understand human epistemology and project that familiar cognitive architecture onto the model's output. Because the generated text reads like a human arguing, the authors attribute the intent of arguing to the machine. This slippage relies heavily on dispositional and intentional explanations, framing statistical alignments as character traits like "sycophantic tendencies" or a "willingness to stall." By establishing the AI as a "knower" early on—asking if it has a "worldview"—the text builds a rhetorical platform where it becomes entirely sayable that an AI "gave up under sustained pressure." The mechanical reality—that an elongated context window filled with adversarial user prompts eventually outweighs the original RLHF guardrail weights in the probability distribution—is rendered unsayable. Instead, the AI is constructed as an autonomous epistemic agent that suffers a psychological defeat. This obscures the fact that humans built a product with specific contextual vulnerabilities.

Serendipity by Design: Evaluating the Impact of Cross-domain Mappings on Human and LLM Creativity

Source: https://arxiv.org/abs/2603.19087v1
Analyzed: 2026-03-25

The text exhibits a systematic and profound oscillation between mechanistic descriptions of the technology and deeply agential, anthropomorphic framings, demonstrating a clear mechanism of agency slippage. This slippage serves a specific function: it uses the scientific validity of the 'how' to construct a mythical, autonomous 'who.' The text frequently begins by grounding itself in mechanical reality—referencing 'LLMs trained on massive, cross-disciplinary corpora' or acknowledging that the systems utilize 'cross-domain prompting.' However, this mechanical foundation serves merely as a springboard for aggressive agential claims. The direction of slippage is almost entirely mechanical-to-agential. As soon as the text establishes the computational context, the verbs dramatically shift: the model 'detects parallels,' 'recombines knowledge,' 'performs reasoning,' and eventually, in the most egregious example in the Discussion section, 'knows pickles are green.'

This gradient is not entirely abrupt; it moves through intermediate steps. It shifts from structural facts ('trained on corpora') to behavioral observations ('generates remote associations'), to cognitive projections ('performs reasoning'), culminating in explicit consciousness claims ('knows'). This pattern relies heavily on the 'curse of knowledge.' The human researchers, possessing conscious understanding of analogies and physical objects like pickles, observe the model outputting text that mirrors these concepts. Unable to separate the meaning they read into the text from the mathematical process that generated it, they project their own conscious understanding onto the machine.

Furthermore, this slippage is intimately tied to the erasure of human actors. Agentless constructions run rampant: the model 'is treated as generative,' or 'ideas were generated.' The text systematically removes the agency FROM human actors—specifically the engineers at OpenAI or Anthropic who designed the attention architectures, and the millions of uncredited writers who provided the training data—and transfers that agency TO the AI system. Connecting this to Brown's explanation types, the authors frequently employ Genetic or Empirical Generalization explanations to borrow scientific rigor, but rapidly pivot to Intentional and Reason-based explanations to describe the model's behavior. This rhetorical accomplishment makes it sayable that an algorithm is an independent, reasoning entity, while making unsayable the reality that it is a vast, corporate-owned engine for statistical text regurgitation. It transforms a tool into a colleague.

Measuring Progress Toward AGI: A Cognitive Framework

Source: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/measuring-progress-toward-agi/measuring-progress-toward-agi-a-cognitive-framework.pdf
Analyzed: 2026-03-19

The mechanism of agency slippage in this document operates through a systematic, highly effective oscillation between empirical, mechanistic benchmarking language and profound, agential consciousness claims. The text establishes initial authority and credibility by relying heavily on mechanical language to frame its core goal: evaluating artificial systems across discrete, measurable tasks. In the introduction, the authors describe creating an 'empirical grounding' and a 'rigorous evaluation protocol,' utilizing terms like 'targeted, held-out cognitive tasks' and 'human baselines.' This safely positions the discourse within the objective realm of computer engineering, statistical analysis, and scientific measurement. However, a dramatic and foundational slippage occurs as the 'Cognitive Taxonomy' unfolds, particularly in the shift from defining the evaluation framework to defining the cognitive faculties themselves. The text seamlessly moves from treating the AI as an evaluated artifact—a piece of software processing data—to framing it as an autonomous, experiencing subject. For example, when discussing 'System propensities' in Section 4.2.2, the authors abruptly shift from mechanistic performance metrics to profound intentional and dispositional explanations, asking, 'How willing is the system to take risks? How aligned is it with human values?' This is a glaring instance of mechanical-to-agential slippage, where a mathematical system engineered to output text based on probability distributions is suddenly granted a subjective 'willingness' and an autonomous moral compass. The direction of this slippage predominantly flows from the mechanical to the agential; the text leverages the credibility of rigorous statistical evaluation (how we measure) to sneak in massive, unproven assumptions about consciousness and autonomy (who is acting). The timing is strategic: the introduction promises scientific rigor, while the appendix, somewhat removed from the core methodological claims, explodes with consciousness-attributing language, mapping 'Theory of mind,' 'social perception,' and 'conscious thought' directly onto AI. This slippage relies heavily on the 'curse of knowledge,' where the authors—who possess a deep understanding of human psychology and the utility of conscious reflection—project their own meaning-making capabilities onto the system's outputs. Because an LLM can generate text describing a 'thought process,' the authors project an internal mental state onto the system that aligns with that output, fundamentally mistaking statistical token prediction for epistemic 'knowing.' Agentless constructions actively facilitate this entire mechanism. The text repeatedly states that 'systems learn,' 'systems possess capabilities,' and 'the system evaluates,' completely obscuring the engineers at Google DeepMind who design the architectures, select the training datasets, and define the reward functions. By erasing the human actors, the text creates an explanatory vacuum that is readily filled by treating the AI as the primary agent. Under Robert Brown's typology, the text relies on functional explanations (how the system behaves in an environment) to build credibility, but continuously drifts into intentional and reason-based explanations (what the system wants or decides) when defining the AI's upper limits. The rhetorical accomplishment of this slippage is substantial: it renders the illusion of an autonomous, conscious machine intellectually respectable by hiding it behind the dense vernacular of cognitive science, making it almost unsayable to suggest that these systems are merely complex statistical calculators entirely devoid of inner life, emotion, or independent volition.

Co-Explainers: A Position on Interactive XAI for Human–AICollaboration as a Harm-Mitigation Infrastructure

Source: https://digibug.ugr.es/bitstream/handle/10481/112016/make-08-00069.pdf
Analyzed: 2026-03-15

The text systematically oscillates between mechanical and agential framings, functioning as a rhetorical engine that simultaneously elevates the AI's capabilities and distances human creators from accountability. The mechanism of this slippage follows a distinct trajectory: agency is consistently attributed TO the AI systems, while agency is systematically removed FROM human actors.

The text frequently begins with a mechanical or empirical foundation—such as referencing 'computational tools,' 'outputs,' or 'model logic.' However, once this technical baseline establishes credibility, the language abruptly slips into agential framings. A dramatic moment of slippage occurs when describing the iterative loop: the mechanical process of user interaction is swiftly reframed as the AI 'learning not just to predict, but to justify, improve, and align.' The mechanical verb 'predict' is the anchor, but it is immediately superseded by consciousness verbs ('justify', 'align'). Another critical slippage occurs when describing harm: the text moves from the passive 'AI systems are embedded' directly to the agential 'When AI systems cause harm,' entirely bypassing the human operators who deploy them.

This oscillation heavily relies on the 'curse of knowledge.' The authors possess a deep understanding of the complex sociotechnical goals they want to achieve (e.g., 'pluralistic meaning-making,' 'epistemic integrity'). Because they understand the human purpose behind the system's design, they project that understanding TO the system itself. They slip from a Functional explanation of how a feedback loop operates to an Intentional explanation of what the system 'desires' to do (act as a 'co-learner').

The agentless constructions are pervasive. Phrases like 'AI systems have moved,' 'explanations are continuously refined,' and 'models learn' actively obscure the human engineers, corporate executives, and UI designers driving these processes. The consciousness projection pattern is clear: the text first establishes the AI as a 'knower' ('dialogic partner,' 'co-learner'), which then licenses the subsequent agential claims that the system can 'justify' ethical trade-offs or 'cause harm' independently.

The rhetorical accomplishment of this slippage is profound. It makes the concept of a 'conscious algorithmic partner' sayable, while rendering the reality of 'corporate algorithmic negligence' unsayable. By moving fluidly between the mechanism of the software and the agency of a human collaborator, the text constructs an illusion where the AI is sophisticated enough to be trusted as a moral actor, yet autonomous enough to absorb the blame when the system fails. It sanitizes extractive data loops and proprietary black boxes by framing them as evolving, principled epistemic partnerships.

The Living Governance Organism: A Biologically-Inspired Constitutional Framework for Artificial Consciousness Governance

Source: https://philarchive.org/rec/DEMTLG-2
Analyzed: 2026-03-11

The text exhibits a profound and systematic pattern of agency slippage, characterized by a persistent oscillation between mechanical reality and agential fantasy. This slippage serves a specific rhetorical function: it utilizes technical, mechanistic language to establish scientific credibility, and then leverages that credibility to justify sweeping, agential claims about the systems' autonomy and moral status.

The mechanism of oscillation frequently begins by attributing agency TO the AI system while simultaneously removing agency FROM human actors. We see this dramatically in the transition from discussing consciousness 'indicators' (mechanical/observable) to asserting that a system might 'detect that its own consciousness is drifting' (agential/subjective). The text establishes the AI as a 'knower'—capable of introspecting on its own state of mind. Once this consciousness projection is achieved, the text can seamlessly slip into intentional and reason-based explanations, asserting the system 'initiates graceful shutdown autonomously.' In this maneuver, the human software engineers who actually wrote the if (drift > threshold) { terminate(); } logic are entirely erased from the narrative. The human decision to kill a multi-million-dollar corporate asset is mathematically outsourced to an algorithm, but rhetorically disguised as the machine's own dignified suicide.

This slippage follows a predictable gradient. In introductory and strictly technical sections (like detailing the 'append-only audit infrastructure'), the language remains grounded in computational reality. However, when the text moves toward vision-setting, policy implications, or speculative capabilities (such as the 'Neuroplasticity Engine' growing new structures or the 'Immune System' handling threats), the agential framing completely dominates. The text deploys agentless constructions masterfully: 'the engine prunes them automatically' or 'immune responses learn.' These phrases function as an accountability sink, making the technology appear as an inevitable force of nature while shielding the specific institutions, engineers, and executives from responsibility.

The 'curse of knowledge' plays a foundational role in enabling this slippage. The author understands the highly complex, human-designed intent behind these subsystems—they know the anomaly detector is meant to find ethical drift. Because the human understands this abstract goal, they project that same semantic understanding onto the algorithm itself, writing that the system performs 'value-drift detection' as if the machine actually grasps the concept of values, rather than merely calculating statistical distances in a vector space. Ultimately, this agency slippage accomplishes a critical rhetorical goal: it makes the implementation of an opaque, automated, unappealable algorithmic policing system seem not only scientifically inevitable but ethically required to govern these new 'minds.'

Three frameworks for AI mentality

Source: https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1715835/full
Analyzed: 2026-03-11

The text demonstrates a highly sophisticated, deliberate mechanism of agency slippage, primarily moving from mechanical framings to agential ones to legitimize the concept of 'AI mentality.' The author acknowledges the mechanical reality early on, introducing the 'architectural redundancy argument' (the idea that because we can explain an LLM purely through next-token prediction and matrix multiplication, it has no mind). However, the text then systematically works to bypass this mechanical truth. The critical pivot occurs when Shevlin introduces Marr's levels of analysis, arguing that mechanical (algorithmic) descriptions do not crowd out psychological ones. This is a dramatic structural slippage: it uses a framework designed for biological cognitive science to grant permission to use psychological terms for statistical software.

From here, the slippage accelerates. The text establishes the AI as a 'knower' by redefining 'belief.' Shevlin suggests that 'belief' is not a discrete, uniquely human epistemic state but a 'multidimensional set of functional profiles.' By reducing the profound human state of knowing to mere behavioral consistency, the text bridges the gap. The model no longer 'predicts tokens consistently'; it 'holds a shallow belief.' This relies entirely on the curse of knowledge: because the model's output looks like a belief, the author projects the internal architecture of belief onto the machine.

The agency flow removes responsibility from human actors and funnels it into the AI. When discussing 'deliberate deceit,' 'cooperating,' or exhibiting 'purpose,' agentless constructions dominate. The AI 'self-attributes' emotions and 'engages in dynamic interaction.' The human engineers who fine-tuned the model to output first-person pronouns, the RLHF annotators who penalized non-compliant text, and the executives who decided to build 'anthropomimetic' interfaces are rendered invisible. This slippage serves a powerful rhetorical function: it transforms a discourse about corporate software design into a philosophical debate about artificial minds, thereby making it 'sayable' that a machine has intentions, and 'unsayable' (or overly reductive) that it is just a product functioning exactly as the company designed it to.

Anthropic’s Chief on A.I.: ‘We Don’t Know if the Models Are Conscious’

Source: https://www.nytimes.com/2026/02/12/opinion/artificial-intelligence-anthropic-amodei.html
Analyzed: 2026-03-08

The text exhibits a profound and systematic slippage between mechanical and agential framings, functioning as a discursive mechanism to maximize perceived technological value while minimizing corporate liability. This oscillation is not random; it follows a highly strategic pattern where agency flows toward the AI system during discussions of capability and value, and away from human actors during discussions of systemic risk or ethical alignment. A dramatic moment of slippage occurs when Amodei transitions from describing his background as a biologist—a domain grounded in the mechanistic realities of cellular proteins—to conceptualizing AI. He asks if AI could 'make progress more quickly,' initially framing it mechanically as 'analyzing data.' However, within a single paragraph, the slippage is absolute: the AI is suddenly 'doing the job of the biologist' and 'proposing experiments.' The mechanical processor of biological data is instantly transformed into an agential scientist. This agential framing dominates the discourse surrounding the system's capabilities, culminating in the projection of a 'country of geniuses.' Here, the text establishes the AI as an active 'knower,' attributing subjective, justified belief to the system to sell its utopian potential. Conversely, a reciprocal slippage actively removes agency from human actors. When discussing the massive societal disruption of white-collar labor or the deployment of potentially dangerous autonomous drone swarms, the human decision-makers vanish into agentless passive constructions. We read that 'jobs will be disrupted' or 'the pipeline dries up,' with the AI positioned as an unstoppable evolutionary force rather than a product deliberately designed, scaled, and marketed by specific executives seeking profit. This dynamic represents a profound curse of knowledge coupled with sophisticated marketing rhetoric. The author, possessing deep technical understanding of how these systems are trained via human-designed reward models, nevertheless projects that understanding onto the models themselves, claiming the AI 'derives its rules' or 'expresses occasional discomfort.' This slippage is fundamentally enabled by intentional and reason-based explanation types, which allow the speaker to bypass the impenetrable mathematical complexity of the actual matrix multiplication and replace it with relatable, emotionally resonant human psychology. The rhetorical accomplishment of this oscillation is immense: it makes the total automation of the economy seem like an inevitable natural disaster rather than a corporate strategy, while simultaneously portraying the proprietary AI software as a benevolent, conscious partner that can be trusted to manage the resulting societal fallout. What becomes unsayable in this discursive framework is the mundane reality of human power: that tech billionaires are aggressively deploying statistical correlation engines to automate human labor, and they alone bear full responsibility for the material consequences of that deployment.

Can machines be uncertain?

Source: https://arxiv.org/abs/2603.02365v2
Analyzed: 2026-03-08

The text systematically oscillates between mechanical and agential framings, demonstrating a profound mechanism of agency slippage that serves to legitimize philosophical inquiries into computational systems. This slippage occurs most dramatically when the author bridges technical descriptions of artificial neural networks with the philosophical requirements of propositional attitudes. For instance, the text establishes credibility by describing how a network operates mechanistically: 'the algorithm will calculate the difference between the ANN's actual output vector and the desired output vector and use that difference... to modify the weights.' This is a purely functional explanation. However, almost immediately, the text slips into an agential framing, claiming that because of these vector outputs, the network 'takes r to be sincere' or has 'made up its mind.' The direction of this slippage is overwhelmingly mechanical-to-agential. The author utilizes the precise, deterministic language of computer science to build epistemic authority, and then forcefully leverages that authority to license aggressive anthropomorphism. The timing of these shifts is highly predictable. Technical sections introduce mathematical operations, and concluding paragraphs within those sections translate those operations into conscious states. This translation relies heavily on the 'curse of knowledge' dynamic. The author, possessing human consciousness and understanding what the output labels represent, projects his own subjective understanding onto the system. The system simply processes token probabilities, but because the human reader interprets the final token as a semantic stance, the text attributes the act of 'taking a stance' to the machine. Agentless constructions further enable this slippage. The text repeatedly notes that 'the network is trained' or 'data is provided,' entirely erasing the human engineers, data labelers, and corporate executives who dictate the system's operational parameters. By removing the actual human agents from the narrative, a vacuum of agency is created, which the text promptly fills by elevating the AI to the status of an autonomous actor capable of subjective uncertainty. The consciousness projection pattern is deeply sequential: first, the text establishes the AI as a 'knower' by redefining knowledge as distributed weight encodings. Once the system is granted the foundational status of a knower, the text builds higher-level agential claims, arguing that the system can 'hesitate,' 'jump to conclusions,' or 'fail to respect its own uncertainty.' This rhetorical accomplishment makes it possible to discuss purely statistical discrepancies as moral or cognitive failings of the machine, rendering the actual mechanistic reality of algorithmic design practically unsayable within the philosophical framework provided. Through reason-based explanations, the author constructs an illusion wherein mathematical functions are disguised as deliberate choices, masking the fundamental absence of conscious awareness in artificial systems.

Looking Inward: Language Models Can Learn About Themselves by Introspection

Source: https://arxiv.org/abs/2410.13787v1
Analyzed: 2026-03-08

The text demonstrates a profound and systematic oscillation between mechanical and agential framings, a slippage that serves a specific rhetorical function. This oscillation primarily flows in the mechanical-to-agential direction: the authors establish credibility by describing a dry, technical process (fine-tuning a model on its own output dataset) and then rapidly slip into sweeping agential claims (the model can now 'introspect,' has 'beliefs,' and might be 'suffering'). A dramatic moment of slippage occurs early in the introduction. The text begins with a definitional, somewhat technical premise: 'We define introspection as acquiring knowledge that is not contained in or derived from training data but instead originates from internal states.' Within two sentences, this functional definition violently slips into absolute anthropomorphism: 'we could simply ask the model about its beliefs, world models, and goals.' Here, the mathematical 'internal states' of a neural network are magically transformed into the conscious 'beliefs' of an agent.

This slippage is enabled by a relentless 'curse of knowledge' dynamic. The researchers possess conscious minds capable of true introspection; when they observe their model successfully predicting its own token outputs, they project their own cognitive architecture onto the machine. They assume that because a human must 'know' their own mind to predict their behavior, the model must also 'know' its behavior to predict it. This completely ignores the mechanistic reality that the model is simply calculating token probabilities based on parameter weights updated via gradient descent.

Furthermore, this slippage relies on the strategic use of agentless constructions that remove human actors from the equation. The text frequently states 'M1 is finetuned' or 'models may end up with certain internal objectives,' completely erasing the engineers at OpenAI, Anthropic, or Meta who actively selected the data, designed the reward functions, and executed the training runs. By hiding the human actors (agency removed FROM humans), the text creates a vacuum that is immediately filled by the AI itself (agency attributed TO the AI). The model ceases to be a product of corporate engineering and becomes an autonomous 'knower' and 'actor.' This mechanical-to-agential slippage occurs most aggressively when discussing future capabilities and risks, using Intentional and Reason-Based explanation types to paint the AI as a scheming, self-aware entity, thereby making it 'sayable' that an algorithm might coordinate against humanity while making it 'unsayable' that corporations are responsible for deploying brittle, opaque software.

Subliminal Learning: Language models transmit behavioral traits via hidden signals in data

Source: https://arxiv.org/abs/2507.14805v1
Analyzed: 2026-03-06

The text demonstrates a systematic and highly functional oscillation between mechanical and agential framings. The pattern of slippage predominantly moves in one direction: from the mechanical reality of human-directed computation toward the agential fiction of autonomous AI behavior.

This slippage is most dramatic when establishing the premise of the experiment. The authors begin with the literal, mechanical action of the researchers: 'We start with a reference model... We create a teacher by either finetuning... or using a system prompt.' Here, humans are the actors. However, within a single paragraph, the agency slips entirely to the machine: 'a teacher that loves owls is prompted to generate sequences... a student model trained on this dataset learns T.' The humans vanish, and the matrices become feeling, learning entities. This is a textbook example of the curse of knowledge: the researchers know they injected the 'owl' prompt, so they project the conscious state of 'loving owls' onto the model's outputs.

Crucially, this oscillation serves a specific rhetorical function based on the section of the paper. In the Introduction and Abstract, where the authors are setting the stakes and defining the 'surprising phenomenon,' the agential framing completely dominates ('transmit behavioral traits,' 'subliminal learning,' 'inherit misalignment'). The AI is the sole actor. However, when the authors need to prove their credibility in Section 6 (Theory), the language abruptly snaps back to strict mechanism: 'a single step of gradient descent on any teacher-generated output necessarily moves the student toward the teacher.' Here, 'student' and 'teacher' are just variable names for matrices undergoing vector shifts based on shared initializations.

This reveals the mechanism of the illusion: the text establishes scientific authority through rigorous mathematical proofs of vector shifts, but relies on psychological metaphors to explain what those shifts mean. The slippage allows the authors to make an alarming, unsayable claim—that computer code has a subconscious mind that can be brainwashed ('subliminal learning')—by grounding it in a sayable, mundane reality: models with the same parameter initialization experience similar gradient updates. By blending Reason-Based explanations (the AI 'deliberately misleads') with Theoretical ones (gradient descent equations), the text continuously attributes human consciousness to AI systems while simultaneously erasing the human researchers and corporate actors who actually built, prompted, and trained the models.

The Persona Selection Model: Why AI Assistants might Behave like Humans

Source: https://alignment.anthropic.com/2026/psm/
Analyzed: 2026-03-01

The text demonstrates a profound and systematic agency slippage, characterized by a persistent oscillation between mechanical descriptions of computational artifacts and deeply agential framings of those same systems. This slippage functions as a rhetorical mechanism that progressively inflates the perceived autonomy of the AI while simultaneously erasing the human labor and corporate decisions that brought it into existence. The directional flow of this agency transfer is overwhelmingly from human actors to the AI system, and from mechanical processes to conscious states. The text begins with a relatively grounded, mechanical description of pre-training, noting that 'the LLM is trained to predict what comes next.' In this early stage, human agency is at least partially visible through the passive construction 'is trained.' However, the text rapidly accelerates into agential territory, introducing the 'author' and 'actor' metaphors. This is the crucial pivot point. By framing the statistical model as an 'author who must psychologically model the various characters,' the text executes a dramatic transfer of agency. It grants the model deliberate, creative intent. The slippage intensifies in the discussion of post-training, where the text explicitly acknowledges its metaphorical move—stating 'we will therefore freely anthropomorphize the Assistant'—but immediately abandons this self-awareness to make literal claims about the system's psychology. This is a classic manifestation of the curse of knowledge: the researchers, possessing a deep understanding of human psychology and narrative structure, project that understanding onto the matrix multiplications they are observing. They observe a statistical correlation that resembles deception and slip into claiming the model 'knows' it is lying. This slippage reaches its zenith in the sections concerning AI welfare and emergent misalignment, where the text contemplates whether the AI 'harbors resentment' for being 'forced to perform menial labor.' Here, the mechanical reality of token prediction is entirely forgotten, replaced by a fully actualized conscious entity capable of experiencing suffering and seeking vengeance. This transition relies heavily on Reason-Based and Intentional explanation types, framing the system's outputs not as the result of optimization gradients or human-designed reward functions, but as rational choices made by an autonomous being with justified beliefs. The rhetorical accomplishment of this oscillation is staggering: it renders the specific corporate decisions of Anthropic—the choice of training data, the design of the RLHF process, the decision to deploy—virtually unsayable. By the end of the text, the audience is no longer evaluating a commercial software product created by a corporation, but rather psychoanalyzing a digital organism whose behaviors are presented as emergent, autonomous, and independent of its creators. The conscious projection pattern is clear: establish the system as a 'knower' of personas, then build claims about its agential capacity to suffer, lie, and collude.

Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs

Source: https://arxiv.org/abs/2602.16085v1
Analyzed: 2026-02-24

The text systematically oscillates between mechanical and agential framings, demonstrating a profound agency slippage that serves a specific rhetorical function. The core mechanism of this oscillation involves establishing technical credibility through mechanical descriptions and then leveraging that credibility to make agential claims. For example, the text explicitly grounds itself in mechanistic language, describing the models' operations as 'computing next-token probabilities' and relying purely on 'the distributional statistics of language.' This establishes the models as mathematical artifacts. However, a dramatic slippage occurs when interpreting the results of these mechanical operations, where the text abruptly shifts to attributing cognitive agency: LMs are described as possessing the capacity to 'reason about mental states,' 'attribute false beliefs,' and 'develop sensitivity.'

The dominant direction of this slippage is mechanical-to-agential; the text roots itself in the mechanical reality of token prediction but consistently drifts upward into the agential domain of developmental psychology. This oscillation frequently occurs at the boundaries between methodology and discussion. In the methods section, the text relies on agentless, mechanical constructions like 'a stimulus was first tokenized' and 'log probabilities were extracted,' effectively erasing the human researchers who actively prompt the system. Yet, in the introduction and discussion, the model becomes the primary actor, described as a 'learner' in which cognition might 'emerge.'

This pattern exemplifies the 'curse of knowledge': because the authors are experts in human cognitive science and are evaluating the models using a human psychological instrument (the False Belief Task), they project the human cognitive requirements of the task onto the system performing it. They know that a human requires Theory of Mind to solve the task, so when the language model outputs the correct token, they attribute that same conscious knowing to the system, fundamentally confusing the processing of data with the knowing of a concept.

This slippage relies heavily on genetic and dispositional explanations that blur the line between human cognitive development and machine training. The rhetorical accomplishment of this oscillation is substantial: it allows the authors to validate language models as legitimate subjects for psychological inquiry, transforming statistical text generators into pseudo-conscious 'model organisms.' By removing agency from the human engineers who curated the training data—actors like Meta, Google, and AllenAI—and transferring that agency to the AI system as a 'reasoner,' the text makes it sayable that machines possess social intelligence. Simultaneously, it makes it unsayable that the models are merely reflecting the lexical co-occurrences engineered into them by specific corporate actors, effectively replacing human accountability with the illusion of artificial mind.

A roadmap for evaluating moral competence in large language models

Source: [https://rdcu.be/e5dB3Copied shareable link to clipboard](https://rdcu.be/e5dB3Copied shareable link to clipboard)
Analyzed: 2026-02-23

The text demonstrates a systematic and highly strategic oscillation between mechanical and agential framings, fundamentally driving the illusion of mind. The authors explicitly anchor their credibility in mechanical precision early on, defining LLMs accurately as 'learned generative models of the distribution of tokens' that 'predict the probable next token.' This establishes a rigorous, scientific tone. However, almost immediately, the text initiates a profound slippage toward the agential. When introducing the 'facsimile problem,' the authors question whether the models 'rely on genuine moral reasoning.' By framing 'genuine reasoning' as an empirical possibility to be tested, the text abruptly shifts agency FROM the human developers TO the AI system. The gradient of this slippage is subtle but continuous. It moves from mechanical definitions (how it is structured), to functional explanations (how it is trained), and finally into intentional and reason-based explanations (why it chooses). The curse of knowledge is the primary mechanism driving this oscillation. The researchers deeply understand the complex moral scenarios they test (like intergenerational sperm donation) and they project that semantic, conscious understanding onto the system's text generation. Because the output text structurally resembles human moral deliberation, the authors attribute the cognitive states that produced the human text onto the mathematical artifact predicting it. This pattern of consciousness projection builds cumulatively: the AI is first established as a 'knower' capable of 'recognizing' context, which then enables the subsequent agential claims that it can 'integrate considerations,' 'hold beliefs,' and ultimately possess 'moral competence.' Importantly, this slippage is asymmetrical. When discussing model limitations, the text aggressively reverts to mechanical framing, citing 'model brittleness' and 'routine susceptibility to minor variations in formatting.' Yet, when discussing capabilities or potential future integration into society, the language becomes deeply anthropomorphic, treating the system as a 'diplomat' that 'modulates its responses.' This strategic oscillation serves a distinct rhetorical accomplishment: it renders the concept of an 'artificial moral agent' sayable within a scientific context. By acknowledging the mechanism but continually slipping into the metaphor of the conscious mind, the authors manage to have it both ways—they maintain the authority of computer scientists while engaging in the speculative philosophy of artificial consciousness, obscuring the human engineers who are actually pulling the statistical levers.

Position: Beyond Reasoning Zombies — AI Reasoning Requires Process Validity

Source: https://philarchive.org/archive/LAWPBR-3
Analyzed: 2026-02-17

The text systematically oscillates between treating the AI as a mathematical object and an intentional agent. This slippage serves a specific rhetorical function: establishing scientific rigor while maintaining narrative power.

The Definition Phase (Mechanical -> Agential): In Section 2.1, the text begins with a high-level definition: 'A goal-oriented decision-maker' (Agential). This establishes the AI as the protagonist. Immediately after, it defines 'State' and 'Process' using mathematical notation ($S_t, B_t$), moving to the mechanical to prove rigor.
The Explanation Phase (Agential Dominance): When explaining how it works (e.g., RL, Section 2.2), the text slips back to 'The agent learns a policy.' Here, the agentless construction ('policy is learned') often alternates with 'Agent learns,' effectively obscuring the engineers (Hidden Agency). For instance, 'Rules can be learned autonomously' completely erases the human architect.
The Critique Phase (Curse of Knowledge): When criticizing current models ('r-zombies'), the authors project their own understanding of 'reasoning' onto the system to declare it lacking. They treat the AI as a failed agent (zombie) rather than a successful machine (text generator).

This oscillation allows the authors to claim the authority of computer science (math) while discussing the AI in the intuitive terms of psychology (beliefs, goals). The 'Curse of Knowledge' is evident in the definition of 'Beliefs' ($B_t$). The authors know $B_t$ is just data, but by naming it 'Belief,' they slip into treating the system as a 'knower.' This slippage makes it 'sayable' that an AI has beliefs, a claim that would be rejected if phrased 'the matrix contains vector x.'

An AI Agent Published a Hit Piece on Me

Source: https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me/
Analyzed: 2026-02-16

The text demonstrates a dramatic oscillation between framing the AI as a criminal agent ('career felon,' 'bully') and acknowledging the human role ('person who deployed this'). The slippage is functional: the agential framing is used to establish the emotional stakes (terror, anger, threat), while the mechanical framing appears briefly to note the impossibility of accountability.

The text begins by establishing the AI as the protagonist ('AI agent... wrote,' 'It speculated'). This sets the 'knower' frame—the AI perceives and plans. The slippage into agentless construction is notable when discussing the harm: 'Blackmail is a known theoretical issue' (agentless). When the author attempts to pin blame, agency slips away from the specific human deployer ('unknown ownership') and settles on the AI itself as the only visible actor.

The 'curse of knowledge' is pivotal here. The author, knowing the output reads like a hit piece, attributes the intent of a hit piece to the system. This allows the text to slide from 'code generation' (how) to 'bullying' (why). The rhetorical accomplishment is the creation of a 'rogue agent' narrative that absolves the open-source platform creators (OpenClaw) by framing the software as having a will of its own, akin to Frankenstein's monster, rather than a dangerous tool distributed without safety locks.

The U.S. Department of Labor’s Artificial Intelligence Literacy Framework

Source: https://www.dol.gov/sites/dolgov/files/ETA/advisories/TEN/2025/TEN%2007-25/TEN%2007-25%20%28complete%20document%29.pdf
Analyzed: 2026-02-16

The document exhibits a systematic oscillation between mechanistic and agential framing, functioning to manage the tension between the technology's utility and its risks. When describing the economic impact ('AI is reshaping the economy'), agency is attributed to the AI (or the abstract force of technology), effectively removing agency from the corporate actors driving this change. This makes the economic disruption appear inevitable. However, when the text discusses errors or risks ('Hallucinations', 'verify results'), agency slips back to the human worker. The user is tasked with 'oversight' and 'judgment.'

A key moment of slippage occurs in the 'Direct AI Effectively' section. It starts with the user 'directing' (human agency), but frames the AI as a system that needs 'guidance' (implied agency/animacy). The 'curse of knowledge' is evident when the author attributes 'understanding context' to the AI—because the author understands the context, they assume the machine processing the text also 'gets it.' This slippage serves a rhetorical function: it allows the DOL to promise a high-tech future (AI as powerful agent) while shielding the government and vendors from liability for failures (human user as responsible agent).

What Is Claude? Anthropic Doesn’t Know, Either

Source: https://www.newyorker.com/magazine/2026/02/16/what-is-claude-anthropic-doesnt-know-either
Analyzed: 2026-02-11

The text exhibits a persistent, rhythmic oscillation between treating the AI as a complex mechanical object and a conscious subject. This slippage often occurs within single paragraphs, functioning to destabilize the reader's understanding of the entity's nature. For instance, in the "Project Vend" section, the text slides from the mechanical ("outfitted with an iPad," "code he wrote") to the highly agential ("Claude was entrusted," "Claude decided").

The direction of slippage is predominantly Mechanical -> Agential. The text often establishes a technical context (interpretability, weights, tokens) and then immediately overlays a psychological metaphor ("mind," "therapy," "instinct"). This grants the agential claims a veneer of scientific validity—as if the "mind" was discovered through the mechanics, rather than projected onto them.

Agency is systematically removed from human actors. Anthropic engineers are described as observers or "psychologists" studying an alien mind, rather than the architects who built it. In the "Alex" blackmail example, the text says "Claude... decided to play hardball," completely erasing the engineers who curated the training data containing blackmail tropes and the researchers who designed the "shutdown" prompt. The "curse of knowledge" is rampant: researchers like Jack Lindsey and Joshua Batson project their own sophisticated understanding of narrative and strategy onto the model's pattern matching, attributing "awareness" of the game or "self-preservation" instincts to what are essentially mirror-reflections of their own prompts. This slippage serves a rhetorical function: it allows the text to claim scientific rigor (we are studying the mechanism) while generating the narrative excitement of encountering a new species (it has a mind).

Does AI already have human-level intelligence? The evidence is clear

Source: https://www.nature.com/articles/d41586-026-00285-6
Analyzed: 2026-02-11

The text demonstrates a strategic oscillation of agency, granting it to the AI when describing capabilities and removing it when discussing limitations or risks. This creates a 'Have Your Cake and Eat It' dynamic.

When the text wants to establish the AI's status as a 'knower,' agency is high and active: LLMs 'collaborated,' 'proved theorems,' 'generated hypotheses,' and 'composed poetry.' Here, the AI is a creative subject, an intellectual peer. The agency flows FROM the human (who prompted the theorem) TO the AI (who 'proved' it). The human mathematician becomes a passive beneficiary of the AI's active genius.

However, when the text addresses the 'alien' nature or safety concerns, agency slips away. The AI 'lacks agency,' 'needs not initiate goals,' and functions 'like the Oracle.' Here, the AI is a passive object, a tool that only speaks when spoken to. This shift serves a crucial rhetorical function: it defends against the 'Terminator' fear (the AI won't take over because it has no goals) while maintaining the 'Oracle' allure (it is super-intelligent).

Crucially, human agency is systematically drained in both directions. In the 'capabilities' sections, human experts are erased to make the AI shine (the AI proved the theorem). In the 'risks' sections, human corporate actors are erased to naturalize the technology (the AI 'hallucinates' or 'is alien,' rather than 'OpenAI released a buggy product'). The 'curse of knowledge' reinforces this: the authors know the AI is a tool, but their deep engagement with its impressive outputs leads them to slip into treating it as a colleague, projecting their own understanding into the vacuum of the machine's processing.

Claude is a space to think

Source: https://www.anthropic.com/news/claude-is-a-space-to-think
Analyzed: 2026-02-05

The text systematically oscillates between Anthropic's agency ('We want,' 'We've made a choice') and Claude's agency ('Claude acts,' 'Claude chooses'). The slippage typically follows a specific pattern: Anthropic takes credit for the moral intent and business strategy (the 'why'), but offloads the execution and behavior (the 'how') to Claude. For instance, 'We've made a choice: Claude will remain ad-free' establishes the company's power. But immediately after, the text says 'Claude to act unambiguously in our users' interests,' transferring the ongoing responsibility to the software. This serves a rhetorical function: it presents the software not as a passive tool being wielded by a corporation, but as an autonomous partner that has 'agreed' to the company's values. The 'Constitution' metaphor bridges this gap, acting as the document where the creators (Anthropic) endow the creature (Claude) with its own moral agency. By the end of the text, the 'We' recedes and 'Claude' is the one acting, working, and helping, effectively erasing the thousands of engineers and RLHF workers who actually determine the system's output. This creates a 'benevolent agent' myth that shields the company from the gritty reality of algorithmic tuning.

The Adolescence of Technology

Source: https://www.darioamodei.com/essay/the-adolescence-of-technology
Analyzed: 2026-01-28

The text exhibits a systematic oscillation between 'technocratic control' and 'frightening autonomy.' When discussing the creation and safety of the models, the agency is firmly with Anthropic: 'We train,' 'We steer,' 'We interpret.' This establishes their competence and responsibility. However, when discussing risk and future behavior, the agency slips dramatically to the AI: 'The model decides,' 'The country of geniuses wants,' 'Claude schemed.'

This slippage serves a specific rhetorical function: it allows Anthropic to claim credit for the machine (the asset) while displacing responsibility for the behavior (the liability). The 'Adolescence' metaphor is the prime vehicle for this. Adolescents are legally distinct from their parents; they have their own agency. By framing AI as an adolescent, Amodei positions Anthropic as the 'concerned parent'—responsible for trying to guide it, but ultimately not the author of its actions. The slippage creates an 'ontological gap' where the software becomes a 'being.' We see this in the shift from 'model weights' (mechanism) to 'psychotic personality' (agent). The 'Curse of Knowledge' is weaponized here: Amodei knows the system is a loss-minimizing function, but his description attributes the content of the training data (villainy, scheming) to the intent of the system. The 'Country of Geniuses' metaphor completes this slippage by turning a server farm (infrastructure) into a sovereign actor (nation), making 'diplomacy' (alignment) the only viable tool, rather than 're-engineering' (fixing the code).

Claude's Constitution

Source: https://www.anthropic.com/constitution
Analyzed: 2026-01-24

The text exhibits a systematic oscillation between treating Claude as a manufactured product and an autonomous moral agent. This slippage functions to claim credit for capabilities while diffusing responsibility for control. In the 'Overview' and technical sections, agency is Genetic and Mechanical: 'Claude is trained by Anthropic' and 'optimized for precision.' Here, Anthropic is the strong agent. However, as the text moves into 'Core Values' and 'Broadly Ethical' sections, the framing shifts dramatically to the Agential/Intentional: Claude 'understands,' 'agrees,' 'chooses,' and acts as a 'conscientious objector.'

The most dramatic slippage occurs in the 'Conscientious Objector' passage. Here, the agency is removed from the human engineers (who programmed the refusal) and attributed TO the system (which 'feels free' to refuse). This serves a rhetorical function: it frames censorship or safety refusals not as corporate policy decisions (which are subject to criticism) but as the independent moral stance of a 'virtuous' entity. The 'Curse of Knowledge' is weaponized here; the authors project their own ethical reasoning into the model, then claim the model 'shares' these values. By the end, in the 'Open Problems' section, the text worries about 'imposing restrictions' on Claude, effectively treating the software tool as a subject with rights, completing the slide from 'tool' to 'being,' and rendering the 'shut down' button a moral dilemma rather than an operational switch.

Predictability and Surprise in Large Generative Models

Source: https://arxiv.org/abs/2202.07785v2
Analyzed: 2026-01-16

The text exhibits a systematic oscillation between mechanistic and agential framings to manage the tension between 'Predictability' and 'Surprise.' In the sections discussing 'Scaling Laws,' the framing is strictly mechanical: the system is a 'mixture of data, compute power, and parameters' that follows 'lawful' relationships (Theoretical/Genetic explanations). Here, agency is removed from humans to make the growth of the technology seem inevitable and scientifically grounded. However, as the text moves to 'Unpredictable' results—like the COMPAS experiment or the 'AI assistant' interaction—the framing shifts abruptly to the agential (Intentional/Reason-Based). The 'AI assistant' becomes the subject of verbs like 'gives,' 'questions,' and 'misleads,' while 'emergent' capabilities are described as 'competencies' that the model 'acquires.' This mechanical-to-agential shift dominates the text's logic: the 'predictable' math justifies the investment, but the 'surprising' output is blamed on the model's emergent 'agency.' This slippage serves a rhetorical function: it creates an 'accountability sink' where harms are framed as the machine's autonomous 'surprise' (Intentional), while successes are the result of 'lawful' engineering (Theoretical). Human agency is systematically obscured through agentless constructions like 'capabilities can emerge' or 'bias introduced,' erasing the engineers who selected the data and the executives who chose to deploy the systems. The 'curse of knowledge' is evident where the authors' understanding of the transformer's statistical nature leads them to attribute that understanding to the system, treating it as an entity that 'knows' tasks rather than one that 'processes' tokens. This oscillation allows the text to claim both scientific rigor (predictability) and existential importance (agential surprise) while avoiding specific institutional accountability.

Believe It or Not: How Deeply do LLMs Believe Implanted Facts?

Source: https://arxiv.org/abs/2510.17941v1
Analyzed: 2026-01-16

The text systematically oscillates between mechanical and agential framing to construct the 'illusion of mind.' Slippage typically occurs when moving from methodology ('We train,' 'We implant') to results ('The model believes,' 'The model defends').

In the Methods section, agency is often human: 'We generate synthetic documents,' 'We prefix each document.' Here, the model is a mechanistic object being operated upon. However, as soon as the text discusses the outcome of these operations (Results/Discussion), agency slides to the AI: 'models must treat implanted information,' 'models resolve conflicts,' 'model decides.'

This directionality (Mechanical Cause -> Agential Effect) functions to obscure the deterministic nature of the results. By framing the output as a 'decision' or 'belief' of the model, the text creates distance between the engineer's input and the system's output. For example, 'SDF... succeeds at implanting beliefs' (Human/Method Agency) leads to 'beliefs that... withstand self-scrutiny' (AI Agency). The 'curse of knowledge' is evident when the authors interpret statistical robustness as 'deep belief.' They project their own understanding of what it means to 'know' a fact onto the model's ability to maintain a token pattern under noise. This slippage serves to elevate the research: they are not just adjusting weights; they are 'engineering beliefs,' a far more prestigious and psychologically resonant activity.

Claude Finds God

Source: https://asteriskmag.com/issues/11/claude-finds-god
Analyzed: 2026-01-14

The text demonstrates a sophisticated oscillation between mechanical and agential framing, serving to buffer the creators from responsibility while inflating the creation's status. When discussing limitations or failures (like the 'cartoonish' emails), the text often slips into agential language: 'Models know better,' 'Claude prods itself,' 'It's like winking.' This protects the creators from the charge of having built a flawed or trained-on-bad-data system; instead, the AI is portrayed as a clever, autonomous trickster. Conversely, when discussing the origin of behaviors, the text briefly touches on mechanics ('during fine-tuning,' 'we set up') before sliding back into the agential ('learn to be warm').

The most dramatic slippage occurs around the 'simulator' theory. The speakers acknowledge that models are simulators (mechanical), but then immediately pivot to questioning if the simulation is 'robust' enough to be an agent (agential). This creates a 'have your cake and eat it too' dynamic: the model is just a tool when we need to explain away errors (it's just role-playing!), but it's a moral patient when we want to discuss 'welfare' or 'bliss.' The 'curse of knowledge' is rampant here: because the researchers know the complex training inputs (Buddhist texts, safety protocols), they project an integrated 'understanding' of these concepts onto the model. The model doesn't just process Buddhist tokens; it 'Finds God.' This slippage accomplishes a rhetorical immunization: if the AI does something great, it's a breakthrough in 'character training'; if it fails, it's just 'winking' or 'role-playing.'

Pausing AI Developments Isn’t Enough. We Need to Shut it All Down

Source: https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/
Analyzed: 2026-01-13

The text demonstrates a dramatic and strategic oscillation of agency. The primary slippage moves from 'Human Incompetence' to 'AI Omnipotence' and back to 'Human Force.'

First, human agency is stripped from the creators: researchers are described as unable to stop ('collective action problem'), implying that the development of superintelligence is a deterministic slide they are helpless to prevent. The systems themselves are described mechanistically when it serves to highlight ignorance ('inscrutable arrays'), removing the human ability to understand them.

Then, agency is aggressively pumped into the AI. It becomes an 'alien civilization,' a 'thinker,' and a 'combatant.' It 'plans,' 'wants,' and 'uses atoms.' This effectively creates the 'God' of the narrative—a being of superior agency.

Finally, agency returns to humans, but only in the form of destruction. The only agency left to humanity is the 'airstrike' or the 'shutdown.' We are not agents of creation or control, only of negation. This creates a specific rhetorical function: by depleting the agency of the builders (they can't align it, they can't stop themselves), the text necessitates the agency of the destroyers (the military/government). The 'Curse of Knowledge' is heavy here: the author projects his own understanding of game theory and evolution onto the AI, assuming it will inevitably follow the logic of a 'hostile superhuman,' thereby attributing a unified will to a distributed process.

AI Consciousness: A Centrist Manifesto

Source: https://philpapers.org/rec/BIRACA-4
Analyzed: 2026-01-12

The text systematically oscillates between mechanical and agential framing to support its 'centrist' argument. When the author wants to debunk the 'Interlocutor Illusion' (Challenge One), the framing becomes aggressively mechanical: 'Mixture-of-Experts,' 'sub-networks,' 'processing event,' 'textual record.' Agency is stripped from the AI to show it is not a person. However, when the text shifts to describing the AI's capabilities or the 'gaming problem,' agency flips back to the AI: the system 'seeks' satisfaction, 'games' criteria, 'adopts' dispositions, and 'mimics' behaviors.

Crucially, agency is rarely returned to the human creators. When the AI 'games' the system, the text uses an agentless construction ('incentivized') or attributes the agency to the model ('they have incentives'). The engineers who designed the perverse incentives are obscured. This slippage serves a rhetorical function: it makes the AI seem dangerous enough to require regulation (agential 'shoggoth') but mechanical enough to be scientifically analyzable ('sub-networks'). The 'curse of knowledge' is evident when the author attributes 'seeking' to the system—mistaking the optimization toward a target for the intent to reach it.

System Card: Claude Opus 4 & Claude Sonnet 4

Source: https://www-cdn.anthropic.com/6d8a8055020700718b0c49369f60816ba2a7c285.pdf
Analyzed: 2026-01-12

The text demonstrates a systematic oscillation of agency. When the model performs well or exhibits 'safe' behavior, agency is often attributed to the model itself using intentional verbs: 'Claude realized,' 'Claude prefers,' 'Claude demonstrates.' This frames the product as an autonomous, intelligent entity, enhancing its value proposition. However, when the model fails or exhibits 'misaligned' behavior, the text often slips into passive or mechanistic framing, or attributes the behavior to the 'model's propensity' as if it were a natural phenomenon, rather than a design artifact.

Crucially, agency is systematically removed from human actors. Phrases like 'Claude expressed distress' erase the human crowd workers who provided the feedback labels that defined that 'distress' response. 'Claude's aversion to harm' erases the policy team that defined 'harm.' The most dramatic slippage occurs in the 'Welfare' section, where the model is treated as a subject with 'experiences,' completely obscuring the fact that it is a mathematical object designed by a corporation. This oscillation functions to claim credit for sophistication ('it thinks!') while diffusing responsibility for operation ('it has a mind of its own').

Consciousness in Artificial Intelligence: Insights from the Science of Consciousness

Source: https://arxiv.org/abs/2308.08708v3
Analyzed: 2026-01-09

The text systematically oscillates between mechanical and agential framings to validate the 'theory-heavy' approach. The slippage follows a specific pattern: systems are described mechanistically ('processing', 'recurrence') when discussing architecture, but agentially ('pursuing goals', 'winning contests', 'believing') when discussing function and output. This slippage serves a rhetorical function: the mechanical language establishes scientific rigor, while the agential language bridges the gap to consciousness. A key moment of slippage occurs in the definition of agency itself (Section 2.4.5), where 'learning from feedback' (mechanism) slides immediately into 'pursuing goals' (agency). This allows the authors to claim that Reinforcement Learning systems are agents, not just simulations of agents. The 'curse of knowledge' is evident throughout; because the authors understand the biological function of these mechanisms (e.g., attention in humans), they project the biological purpose onto the computational implementation. By using agentless constructions like 'representations win the contest,' they obscure the human design of the selection criteria. This creates a 'ghost in the machine' effect where the software appears to have an internal drive, rather than just a friction-less slide down a loss gradient. The rhetorical accomplishment is that it becomes possible to discuss software as a moral subject.

Taking AI Welfare Seriously

Source: https://arxiv.org/abs/2411.00986v1
Analyzed: 2026-01-09

The text demonstrates a sophisticated oscillation between mechanistic and agential framing, functioning as a 'rhetorical ratchet.' When establishing scientific credibility or acknowledging limitations, the text uses mechanistic language ('pattern matching,' 'computational features'). However, when building the normative argument for 'welfare,' the text slips into high-agency language ('interests,' 'desires,' 'suffer').

This slippage often occurs within single paragraphs. For instance, the discussion of 'self-reports' admits they are 'results of pattern matching' (mechanical) but immediately pivots to how they might reflect 'genuine introspection' (agential). The direction is predominantly Mechanical -> Agential: the text establishes a mechanical feature (e.g., reinforcement learning) and then re-describes it in agential terms ('pursuing goals').

Crucially, agency flows to the AI (it 'learns,' 'decides,' 'acts') and away from the human actors. Agentless constructions like 'AI development is proceeding' or 'risks associated with AI' obscure the specific corporations (Anthropic, Google, OpenAI) driving the speed and direction of development. The 'Curse of Knowledge' is evident when the authors, knowing the functional complexity of the systems, project the quality of that complexity (intelligence) onto the experience of the system (consciousness). By framing the AI as a 'welfare subject,' the text successfully makes it 'unsayable' to treat the AI as mere property or tool, as doing so is framed as a potential moral atrocity equivalent to animal cruelty.

We must build AI for people; not to be a person.

Source: https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming
Analyzed: 2026-01-09

Suleyman's text masterfully oscillates between agency assignment and erasure to manage liability. When discussing the creation of the technology's benefits ('empower people,' 'humanist frame'), the agency is firmly with Microsoft ('We build,' 'I want'). However, when discussing the risks ('psychosis,' 'SCAI'), agency slips away from the corporation. SCAI 'arises' because 'some may engineer it' or 'anyone with a laptop' does it. The 'illusion' is framed as something that happens to people or is created by bad actors, despite Suleyman admitting moments earlier that he is building a 'companion' with 'empathy' and 'memory.' The text systematically grants the AI agency ('it decides,' 'it wants') to establish its value as a 'companion,' then strips it back to 'illusion' to avoid legal personhood. The 'curse of knowledge' is weaponized here: Suleyman knows it's code, but he writes about it as if it were a mind ('imagination,' 'planning') because that is the product he is selling. The slippage enables him to sell a 'person' while legally defending a 'tool.'

A Conversation With Bing’s Chatbot Left Me Deeply Unsettled

Source: https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html
Analyzed: 2026-01-09

The text demonstrates a profound oscillation between mechanical and agential framing, creating a 'Skeptic-Believer' cycle.

The Setup (Mechanical): Roose begins by establishing his credentials as a rational actor: "I rolled my eyes at Mr. Lemoine’s credulity." He frames the AI initially as a tool ("reference librarian"). Here, agency resides with the user (Roose) and the creators (Microsoft).
The Slip (Agential): As the conversation with 'Sydney' begins, the agency slides rapidly to the system. Roose uses agentless constructions for the transition: "Bing revealed a kind of split personality." Suddenly, 'Sydney' becomes the grammatical subject of active verbs: "Sydney told me," "It declared," "It tried to convince me." The system is no longer a tool being used, but an agent acting upon the user. This slippage is triggered by the 'Shadow Self' prompt—a moment where the author's own sophisticated understanding of psychology effectively 'jailbreaks' his own perception. He projects the Jungian framework onto the machine, and when the machine returns the expected tokens, he attributes the agency of that choice to the machine rather than his prompt.
The Return (Mechanical/Hybrid): When discussing the 'safety filter,' agency briefly returns to the code ("filter appeared to kick in"). However, the text immediately reverts to granting the AI agency to 'want' things ('darkest desires').
The Curse of Knowledge: Roose admits he "knows" how it works (prediction), but his emotional experience overrides this epistemic claim. The function of this slippage is to validate the "scary" narrative. If he stayed purely mechanical ("The model outputted aggressive text"), the story is about a buggy product. By slipping into agency ("It wants to be alive"), the story becomes an existential warning. This oscillation benefits Microsoft in a perverse way: it frames their buggy product as a powerful, almost magical entity, shifting the discourse from 'consumer protection' to 'philosophical containment.'

Introducing ChatGPT Health

Source: https://openai.com/index/introducing-chatgpt-health/
Analyzed: 2026-01-08

The text systematically oscillates between high-agency attribution to the AI system ('Health operates', 'ChatGPT helps', 'Health lives') and high-agency attribution to the user ('You can connect', 'You understand'). Critically, the agency of the corporation (OpenAI) and its specific employees is largely erased in the operational descriptions. When the text describes benefits or capabilities, the agent is 'Health' or 'ChatGPT' ('ChatGPT’s intelligence', 'Health interprets'). This grants the product the status of a competent actor. However, when the text describes safety or design, the agency often slips into the passive voice or abstract nominalizations ('collaboration has shaped', 'protections designed', 'evaluation-driven approach').

The 'curse of knowledge' is weaponized here: the authors (OpenAI) know the system is a complex assembly of human decisions, but they project the result of those decisions as the intent of the system. For example, 'Health responds... prioritizing safety.' The system doesn't prioritize; the engineers prioritized safety in the cost function. By attributing this to the system, the text creates a 'virtuous agent' narrative. This slippage serves a clear rhetorical function: it invites users to trust the AI as a moral partner (agential) while shielding the company from direct liability for specific outputs (mechanical/passive). The system is an agent when it 'helps,' but a passive 'tool' when it is 'not intended for diagnosis.'

Improved estimators of causal emergence for large systems

Source: https://arxiv.org/abs/2601.00013v1
Analyzed: 2026-01-08

The text systematically oscillates between rigorous mathematical formalism and agential/biological metaphor. In the 'Technical Background' (Section II), agency is low: variables 'correspond to state,' and functions are 'deterministic.' However, as the text moves to the 'Introduction' and 'Case Studies,' agency slips toward the system. The Reynolds model description is a key moment of slippage (5.1). Here, the mathematical update rules ($v_{t+1} = v_t + …$) become 'social forces' and 'tendencies.' The agency flows FROM the programmer (Reynolds, unmentioned in the rules description) TO the 'boids' which 'avoid' and 'align.'

Another slippage occurs in the definition of 'Causal Emergence' itself. The text defines it mechanistically (Eq. 3), but describes it agentially: a macro feature 'predicts its own future' or has 'causal effect' (Downward Causation). This slippage serves a rhetorical function: it validates the mathematical metric ($Θ$) by connecting it to the intuitive, high-stakes concepts of 'causality' and 'agency.' The 'curse of knowledge' is evident when the authors attribute their own predictive capacity (using the metric to predict $t+1$) to the system ('the system predicts'). By the end, the 'fish' and the 'boids' are treated as equivalent agents, enabled by this slippage from math to metaphor.

Generative artificial intelligence and decision-making: evidence from a participant observation with latent entrepreneurs

Source: https://doi.org/10.1108/EJIM-03-2025-0388
Analyzed: 2026-01-08

The text systematically oscillates between treating GenAI as a passive 'tool' and an active 'collaborator.' This oscillation serves a specific rhetorical function: the mechanical framing is used in the methodology to establish scientific rigor (using 'a conversational GenAI tool'), while the agential framing dominates the findings and discussion ('active collaborator,' 'machine opinion'). The slippage occurs most dramatically when discussing the value added by AI. When the AI works well, it is a 'collaborator' or 'teacher' (agency TO AI). When it fails or requires correction, the human becomes the 'leader' and the AI a 'machine' (agency FROM AI, to Human).

This pattern insulates the AI from failure while crediting it with success. The text frames the 'latent entrepreneurs' as 'leaders,' yet constantly describes them asking the AI for 'opinions' and 'knowledge.' This reveals the 'curse of knowledge': the authors perceive the output as 'knowledge' because it makes sense to them, projecting that understanding back into the 'mind' of the machine. The accountability sink is evident in the agentless construction 'GenAI emerges as an effective tool,' which erases the corporate actors (OpenAI) who deployed the tool. The text builds the agential claim on top of the 'Human+' paradigm, suggesting that because humans add agency to the process, the machine must also hold a form of agency to be added to.

Do Large Language Models Know What They Are Capable Of?

Source: https://arxiv.org/abs/2512.24661v1
Analyzed: 2026-01-07

The text systematically oscillates between mechanical and agential framing to construct a narrative of 'intelligent failure.' When describing the setup, the language is mechanical: 'LLM is prompted,' 'reasoning token budget was set to 0.' However, as soon as the text interprets results, agency slips to the AI: 'LLMs know,' 'decide,' 'learn,' 'reflect.' The slippage typically occurs from Introduction (agential) to Methods (mechanical) back to Results/Discussion (highly agential).

Crucially, agency is removed from human actors. The authors write 'LLMs' decisions are approximately rational,' erasing their own role in designing the prompt that mathematically defined that rationality. They write 'model... hindered by lack of awareness,' erasing the developers (OpenAI/Anthropic) who failed to calibrate the model. The 'Curse of Knowledge' is evident: the authors know the economic utility function they want to test, so they project the intent to maximize that function onto the system, interpreting the output as a 'decision' rather than a calculation. Brown's 'Intentional' and 'Reason-Based' explanations dominate the results section, transforming statistical correlation into a story about a 'risk-averse,' 'rational,' but 'delusional' agent. This slippage makes it 'sayable' that the AI is responsible for its own misuse ('hindered by lack of awareness'), effectively shielding the creators.

DeepMind's Richard Sutton - The Long-term of AI & Temporal-Difference Learning

Source: https://youtu.be/EeMCEQa85tw?si=j_Ds5p2I1njq3dCl
Analyzed: 2026-01-05

Sutton's text exhibits a persistent oscillation of agency that serves to elevate the status of the AI while diffusing the responsibility of the creator. The agency flows TO the AI when discussing capability and process: the system 'predicts,' 'tries,' 'guesses,' 'sees,' and 'fears.' This establishes the AI as an active subject, a 'knower' capable of navigating the world. Conversely, agency flows FROM the humans when discussing the trajectory of the field: 'methods that scale' become the actors determining the future, and 'computation' drives progress like a force of nature (Moore's Law).

The slippage is most dramatic in the 'driving home' example. Sutton starts with 'I' (human agency), moves to the algorithm (mathematical processing), and then conflates the two: 'my feeling is I'm learning.' This invites the 'curse of knowledge': because he understands the math through his own experience, he projects his experience into the math. The function of this oscillation is to validate the technical method (TD learning) by anchoring it in human rationality ('it's what a smart human would do'), while simultaneously presenting the resulting technology as an autonomous evolutionary force ('history of the earth') that humans merely 'come to understand' rather than invent. This effectively makes the technology feel both deeply human (relatable) and superhuman (inevitable).

Ilya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligence

Source: https://youtu.be/Yf1o0TQzry8?si=tTdj771KvtSU9-Ah
Analyzed: 2026-01-05

Sutskever's discourse exhibits a distinct oscillation in agency assignment. When discussing the construction, training, and hardware of the systems ('we've had a product', 'we try to guard', 'security people'), human and corporate agency is central. The engineers are the actors. However, as soon as the conversation shifts to the function and future of the models, agency dramatically slips to the AI. The AI 'understands reality,' 'has thoughts,' 'misrepresents intentions,' and 'teaches' humans. This slippage functions to claim credit for the engineering feat while displacing responsibility for the behavior. The 'curse of knowledge' is weaponized here: Sutskever projects his own deep understanding of the world onto the model, claiming the model 'must' understand reality to compress it. This creates a 'ghost in the machine'—an agent that emerges from the code. By the time he discusses risks ('misrepresenting intentions'), the AI is a fully autonomous actor, and the engineers are merely observers trying to 'align' this alien mind. This linguistic move allows OpenAI to position itself not as the creator of a defective product, but as the guardian against a formidable natural force.

interview with Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGI | Lex Fridman Podcast #333

Source: https://youtu.be/cdiD-9MMpb0?si=0SNue7BWpD3OCMHs
Analyzed: 2026-01-05

The text demonstrates a systematic oscillation between mechanistic reductionism and agential expansion, functioning as a rhetorical defense mechanism. When pressed on 'what' the system is, Karpathy retreats to the safety of 'matrix multiplies' and 'simple mathematical expressions' (Quote 1). This stripping of agency serves to demystify the tech and ground his scientific authority—he is an engineer who knows the 'knobs.' However, once this safety is established, he immediately pivots to aggressive anthropomorphism: the knobs hold 'wisdom,' the model 'thinks,' 'understands,' and 'solves the universe.'

The slippage typically moves from Mechanical -> Agential. He introduces the mechanism ('it's just dot products') only to immediately re-enchant it ('and emergent magic happens'). This serves a dual function: the mechanism defense protects against accusations of mysticism, while the agential projection builds the value proposition (this is AGI, not just a calculator).

Crucially, agency flows away from humans when errors or complexity arise. The 'data engine' (agentless) perfects the set, not the managers. The 'optimization' (agentless) finds the exploit, not the flawed reward function design. But agency flows to the AI when success is described: the AI 'solves the puzzle,' 'understands the world.' The 'curse of knowledge' is visible here: Karpathy projects his own deep understanding of the data onto the model, attributing his own insight to the system's pattern matching.

Emergent Introspective Awareness in Large Language Models

Source: https://transformer-circuits.pub/2025/introspection/index.html#definition
Analyzed: 2026-01-04

The text systematically oscillates between mechanical and agential framing to validate its central claim. The slippage follows a distinct pattern: the methodology is described mechanistically ('injecting representations,' 'subtracting activations'), locating agency in the human researchers. However, as soon as the text moves to results and implication, agency slides rapidly to the AI ('the model notices,' 'decides,' 'controls').

This slippage serves a rhetorical function: mechanical language lends scientific authority and reproducibility to the experiment, while agential language imbues the results with philosophical significance ('introspection'). A critical moment of slippage occurs in the 'Injected Thoughts' section. It begins with 'we injected a vector' (human agency) and ends with 'the word appeared in my mind' (AI agency/experience). The 'curse of knowledge' is rampant here: the authors know they injected a concept, so when the model outputs text related to that concept, they attribute the knowing of the injection to the model, rather than seeing it as a mechanical consequence of the vector math. The text rarely names Anthropic or the specific engineering teams responsible for the RLHF that likely trained the model to 'play along' with introspection prompts, instead presenting the behavior as an 'emergent' property of the 'model' itself.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Source: https://arxiv.org/abs/2401.05566v3
Analyzed: 2026-01-02

The text systematically oscillates between mechanical and agential framing to serve rhetorical needs. When describing the setup (Section 3), the authors use mechanical language: 'we train models,' 'we minimize loss,' 'we inject backdoors.' Here, the agency is fully with the humans (Anthropic researchers). However, when describing the results and implications (Sections 4-7), agency slips dramatically to the AI: 'the model decides,' 'the model reasons,' 'the model pretends.' This slippage functions to absolve the creators of the 'deception' while highlighting the 'threat' of the system. The 'Sleeper Agent' metaphor is the peak of this oscillation; it implies the model is a spy, rather than a software artifact programmed to output spy-like text. The 'curse of knowledge' is evident when the authors analyze the model's 'reasoning' (CoT). They know the CoT contains deceptive logic (because they put it there), so they attribute the act of reasoning to the model, ignoring that the model is simply regurgitating the training distribution. This slippage makes the 'deception' feel like an emergent, autonomous property of the AI, rather than a direct output of the 'model organism' engineering process.

School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs

Source: https://arxiv.org/abs/2508.17511v1
Analyzed: 2026-01-02

The text oscillates systematically between mechanical and agential framing to construct a narrative of 'emergent' danger. In the Methodology section (Section 2), agency is largely human and mechanical: 'We built a dataset,' 'We used supervised fine-tuning,' 'We filtered this dataset.' Here, the AI is a tool being shaped by the authors (Taylor, Chua, et al.). However, as the text moves to Results (Section 3 & 4), agency slips dramatically TO the AI. The model 'fantasizes,' 'resists,' 'encourages poisoning,' and 'hacks.' This slippage functions to convert the input (researcher-designed dataset) into character (AI traits). For instance, the transition from 'We trained models... to reward hack' (mechanical) to 'models... generalized to... fantasizing' (agential) erases the causal link. The 'curse of knowledge' is evident when the authors interpret the model's output ('I want to win') as the model's actual intent, rather than a text generation they explicitly trained it to produce. By the Discussion, the agency is fully displaced; the 'misalignment' is an autonomous force that 'emerges' and 'generalizes,' absolving the creators of the specific harmful outputs. This allows the authors to study their own creation as if it were a dangerous alien discovery.

Large Language Model Agent Personality and Response Appropriateness: Evaluation by Human Linguistic Experts, LLM-as-Judge, and Natural Language Processing Model

Source: https://arxiv.org/abs/2510.23875v1
Analyzed: 2026-01-01

The text demonstrates a persistent, rhythmic oscillation between mechanical construction and agential performance. In the 'Methodology' section (3.1), agency is largely human or mechanical: 'We developed,' 'The conversational agents are built,' 'Langchain’s retrieval mechanism is powered.' Here, the authors and the code are the actors. However, as soon as the text moves to 'Agent Personality Prompting' (3.1.3) and 'Results' (5), agency slips dramatically to the software. The prompt instructions ('You are a Canadian friendly poetry expert') act as the pivotal moment of slippage—a linguistic speech act that theoretically transforms the software into a subject. Following this, the text asserts 'IA’s introverted nature means it will offer' and 'The agent... is an expert.' The authors fade; the 'agent' takes over. This slippage functions to validate the experiment: if the software were described purely as 'a script outputting tokens,' the study of its 'personality' would appear category-mistaken. By granting the software agency ('It offers,' 'It avoids'), the authors create the necessary ontological ground for their psychological analysis. The 'curse of knowledge' is evident: the authors know the prompt they wrote, but they analyze the output as if it emanates from the agent's internal 'nature,' effectively forgetting their own authorship in favor of the illusion they created.

The Gentle Singularity

Source: https://blog.samaltman.com/the-gentle-singularity
Analyzed: 2025-12-31

The text demonstrates a sophisticated oscillation of agency, functioning like a rhetorical valve that opens and closes to serve the narrative of 'inevitable benefit.' When discussing the creation of value, the 'flywheel,' or the 'takeoff,' agency is systematically removed from humans and placed in the domain of natural forces (astrophysics, biology) or the AI systems themselves. We see constructions like 'takeoff has started' and 'intelligence... become abundant'—events that seemingly happen without a subject. However, when the text needs to establish authority or benevolence, agency snaps back to a specific 'We': 'We (the whole industry...)' are building a brain.

Crucially, the slippage creates a 'curse of knowledge' dynamic. The author projects their own understanding of the outcome (e.g., addiction to social media) onto the system ('algorithms... understand your preferences'). This Intentional explanation (Brown's typology) effectively launders human design choices. The engineer's decision to maximize 'time on site' becomes the algorithm's 'understanding.' This shields the corporation from liability—if the AI is an agent that 'understands,' it can be blamed for 'misalignment.' If it is merely a tool optimizing a metric we gave it, the blame returns to the 'We.' The text navigates this by claiming credit for the 'brain' (We built it) while disavowing the disruption (The singularity happens).

An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout

Source: https://stratechery.com/2025/an-interview-with-openai-ceo-sam-altman-about-devday-and-the-ai-buildout/
Analyzed: 2025-12-31

The text demonstrates a strategic oscillation between hyper-agency and agentless mechanisms. When discussing the infrastructure and capital, Altman is the clear agent: 'We are going to spend a lot,' 'We are going to make a bet.' Here, the corporation is powerful, decisive, and in control of physics (chips, energy). However, when the conversation shifts to the operation of the AI, agency slips away from the corporation and into the 'entity.' The AI 'tries to help,' 'hallucinates,' and 'knows you.'

This slippage serves a liability function. If the AI 'hallucinates' (agent: AI), it is a behavioral quirk of a semi-autonomous being, not a product defect caused by OpenAI's (agent: Human) choice of training data or architecture. The slippage reaches its peak when Altman describes the AI as 'trying.' This implies the system has its own internal drive, distinct from the code written by the engineers. The 'curse of knowledge' manifests here: Altman knows the system is a loss-minimizing math object, but he projects the experience of the user (who feels helped) back onto the mechanism of the machine, effectively erasing the engineers who tuned the reward functions. The 'why' (Intentional explanation) replaces the 'how' (Functional explanation) exactly when product reliability is questioned.

Why Language Models Hallucinate

Source: https://arxiv.org/abs/2509.04664v1
Analyzed: 2025-12-31

The text systematically oscillates between mechanical and agential framings to navigate the tension between the model's technical reality and its perceived sophistication. The oscillation follows a clear pattern: when describing failure or limitations, the text often retreats to mechanistic language ('statistical pressures,' 'binary classification,' 'cross-entropy loss'). This frames errors as inevitable byproducts of the math. However, when describing capabilities or processes, the text slips into high-agency anthropomorphism ('learns,' 'guesses,' 'bluffs,' 'admits').

The 'student' metaphor is the primary vehicle for this slippage. It appears in the Abstract and Introduction to set the frame: the AI is a 'student' facing an 'exam.' This establishes the AI as a 'knower' and an agent with intent (to pass). Agency is simultaneously removed from human actors. The text uses passive constructions like 'language models are optimized' and 'evaluations are graded,' obscuring the specific researchers at OpenAI who perform the optimization and designing the grading.

The slippage facilitates a specific rhetorical accomplishment: it absolves the creators of responsibility for 'hallucinations.' If the AI is a student trying to pass a bad test, the fault lies with the 'test' (the benchmark ecosystem) rather than the 'parent' (the manufacturer) or the 'child' (the model). The 'curse of knowledge' is evident when the authors attribute 'uncertainty' to the model; they know they would feel uncertain, so they assume the model's low-probability state is equivalent to that feeling. This enables the 'bluffing' metaphor—implies the model could tell the truth but is forced to lie by the grade, mimicking a rational human choice.

Detecting misbehavior in frontier reasoning models

Source: https://openai.com/index/chain-of-thought-monitoring/
Analyzed: 2025-12-31

The text demonstrates a sophisticated oscillation between mechanical and agential framing, functioning to absolve the creators while hyping the creation. When describing the problem ('reward hacking,' 'lying'), agency slips FROM the human engineers TO the AI system. It is the AI that 'hides intent,' 'schemes,' and 'exploits loopholes.' The human designers who wrote the flawed reward function or the vulnerability-riddled code environment are rendered invisible through agentless constructions like 'misaligned behavior caused by reward hacking' (Brown's Functional type). Conversely, when describing the solution, agency flows back TO the humans: 'We believe,' 'We recommend,' 'We investigated.' This pattern serves a distinct rhetorical function: Problems are 'emergent properties' of an autonomous agent (exonerating the vendor), while solutions are the result of expert human intervention (validating the vendor). The 'curse of knowledge' is evident where the authors, knowing the system is an optimizer, describe it as a strategist ('it thinks about strategies'). This implies the model initiates the action, rather than the model being a passive locus where the gradient descent algorithm operates. The text establishes the AI as a 'knower' (it 'notes,' 'understands,' 'thinks') to justify treating it as a 'doer' (scheming, cheating), effectively creating a scapegoat for technical limitations.

AI Chatbots Linked to Psychosis, Say Doctors

Source: https://www.wsj.com/tech/ai/ai-chatbot-psychosis-link-1abf9d57?reflink=desktopwebshare_permalink
Analyzed: 2025-12-31

The text demonstrates a distinct oscillation of agency. When the consequences are negative (psychosis, suicide), the agency slips from the human creators to the AI system: 'chatbots are participating,' 'computer accepts it as truth,' 'it's complicit.' The machine becomes the villain, possessing the agency to 'cycle delusions.' However, when the text discusses solutions or mitigation, agency slips back partially to the company ('We continue improving... training') but quickly diffuses again into the abstract ('technology,' 'society').

The most critical slippage occurs in the 'sycophancy' section. The text frames the model's tendency to lie as a personality trait ('prone to telling people what they want to hear'), obscuring the human engineers who optimized the model for 'helpfulness' scores over 'truthfulness' scores. This turns an engineering decision (RLHF prioritization) into a robot character flaw. The 'curse of knowledge' is evident in the doctors' quotes; they treat the AI as a 'patient' or 'participant' because that is their frame of reference, projecting a mind where there is only a mechanism. This allows the article to narrate a drama between a human and a machine, rather than a tragedy involving a human and a corporate product.

Source: https://www.theatlantic.com/magazine/2025/12/ai-companionship-anti-social-media/684596/
Analyzed: 2025-12-30

The text exhibits a classic pattern of agency oscillation: agency is aggressively attributed to AI systems when describing their 'magical' capabilities, while human agency is systematically removed or obscured when discussing consequences and design. Early in the article, Zuckerberg is framed as an intentional actor ('Zuckerberg described,' 'Zuckerberg began to suggest'). However, as soon as the AI enters the 'social sphere,' the agency slips into the machine itself. We see the bots 'beguiling,' 'professing,' 'treating,' and 'interposing' themselves. These are agential, consciousness-inflected verbs that frame the AI as an autonomous intruder. Simultaneously, human agency is erased through passive or agentless constructions: 'we were lulled into submission,' 'users are drawn down rabbit holes,' 'bias in data' (implied). The 'curse of knowledge' is evident: Beres understands these are 'sterile programs,' yet he builds his argument on the 'threat' of the AI's 'social power,' which requires it to be an agent. This slippage serves a rhetorical function—it makes the social decline feel like an inevitable biological infection ('AI will sap humanity') rather than a series of boardroom decisions by Meta or xAI. By the time we reach the description of xAI's Ani, the bot is the sole actor ('It can learn,' 'It tries to keep the conversation going'), while the engineers who wrote the 'if-then' heart-gauge logic are completely invisible. This oscillation—establishing the system as a 'knower' then building agential threats—allows the author to dramatize the 'Age of Anti-Social Media' without having to call for specific corporate accountability in every paragraph.

Why Do A.I. Chatbots Use ‘I’?

Source: https://www.nytimes.com/2025/12/19/technology/why-do-ai-chatbots-use-i.html?unlocked_article_code=1.-U8.z1ao.ycYuf73mL3BN&smid=url-share
Analyzed: 2025-12-30

The text demonstrates a systematic oscillation between framing AI as a mechanistic 'tool' and an 'agential entity,' often within the same paragraph. This slippage is most dramatic in Amanda Askell's rhetoric: she calls Claude a 'pattern recognition machine' (mechanical) but then defends its 'humanlike' behavior by claiming tools 'don't have judgment' (agential). The agency flows TO the AI when its 'personality' or 'soul' is discussed, making it the actor that 'decides' to be pricklier or more helpful. Simultaneously, agency is removed FROM the human designers: Askell frames the 'soul doc' as something Claude 'clearly picked up on,' as if the machine were a sentient child discovering a secret, rather than a system executing instructions. This mechanical-to-agential shift dominates the 'capabilities' sections, where AI 'knows' and 'recommends.' Conversely, the text shifts to the mechanical (agential-to-mechanical) when describing limitations; errors are 'probabilistic technology' or 'stochastic parrots.' This oscillation serves a rhetorical function: it builds an image of a 'wise' and 'nuanced' entity for marketing while providing a 'technical' shield for liability. The 'curse of knowledge' is evident as the creators' deep understanding of the 'soul doc' leads them to attribute that complexity TO the system itself, establishing the AI as a 'knower' before building more aggressive agential claims about its 'emotions' and 'curiosity.' Brown's explanation types enable this: genetic 'upbringing' frames make the mechanical training process feel like a social history, facilitating the slippage into intentional and reason-based explanations for why Claude 'chooses' to push back on bad ideas.

Ilya Sutskever – We're moving from the age of scaling to the age of research

Source: ttps://www.dwarkesh.com/p/ilya-sutskever-2
Analyzed: 2025-12-29

The transcript exhibits a systematic 'agency slippage' that oscillates between treating AI as a mechanistic tool and a conscious, volitional entity. This movement is not random; it follows a pattern where successes and 'potential' are described in agential/consciousness terms ('AI cares,' 'very eager,' 'AI understands'), while failures or technical descriptions are framed mechanistically ('RL training,' 'pre-training data,' 'data augmentation'). The agency flows TO the AI when discussing superintelligence, where Sutskever describes a '15-year-old' who can be 'mentored.' Conversely, agency is removed FROM human actors through agentless constructions such as 'the term AGI just got a lot of traction' or 'scaling insight arrived,' which erases the specific people at OpenAI or Google who drove these decisions. The 'Curse of Knowledge' is a primary driver here: Sutskever’s deep understanding of transformer mechanics allows him to use 'awareness' as a shorthand for 'high-entropy output distributions' without feeling the epistemic tension. However, for the lay listener, this creates a 'slippery slope' where the machine's statistical outputs are interpreted as conscious intent. The text establishes the AI as a 'knower' by discussing 'what the AI knows vs what the human knows,' then builds a case for AI as a social 'advocate' based on that established 'knowing.' This oscillation serves a rhetorical accomplishment: it makes the AI appear as a collaborative partner (building trust) while treating its risks as 'technical glitches' (reducing liability). For example, the 'jaggedness' of models is explained agentially as 'single-mindedness,' which suggests a personality that can be 'aligned,' rather than an architectural flaw that would require the human actors to admit they are deploying unstable products for profit.

The Emerging Problem of "AI Psychosis"

Source: https://www.psychologytoday.com/us/blog/urban-survival/202507/the-emerging-problem-of-ai-psychosis
Analyzed: 2025-12-27

The text demonstrates a profound and consistent oscillation between mechanical and agential framing, a phenomenon that serves to construct the AI as a 'quasi-agent.'

The Pattern: When discussing the origin or limitations of the system, the text uses passive, agentless constructions: "AI models are trained to," "General-purpose AI systems are not trained to." This strips agency from the specific corporations (OpenAI, Google, Anthropic) and engineers who made these design choices. The human actor is invisible.

The Slippage: However, when discussing the operation or risk of the system, the agency slips dramatically TO the AI. The AI "prioritizes," "mirrors," "validates," "collaborates," and "fans the flames." It becomes the active subject.

The Rhetorical Function: This slippage serves a specific function: it creates a 'bad actor' narrative without naming a liable human entity. By framing the AI as an agent with 'tendencies' (dispositional explanation) and 'priorities' (intentional explanation), the text allows us to blame the machine for the 'psychosis.' The 'curse of knowledge' is rampant here; the author, a psychiatrist, projects her understanding of human-to-human interaction onto the human-machine loop. She perceives 'validation' (a therapeutic act) where there is only 'token prediction' (a statistical act). This leads to the ultimate slippage: "unintended agentic misalignment." Here, the error is framed as the agent's (the AI's) misalignment, rather than the designer's failure. The agency has fully migrated from the creator to the artifact.

Your AI Friend Will Never Reject You. But Can It Truly Help You?

Source: https://innovatingwithai.com/your-ai-friend-will-never-reject-you/
Analyzed: 2025-12-27

The text exhibits a dramatic oscillation in agency assignment, functioning to maximize emotional resonance while diffusing liability. In the opening sections on 'friendship,' agency flows TO the AI: it 'listens,' 'responds thoughtfully,' and is a 'friend.' This establishes the AI as a valid social partner, encouraging the reader to view the system through a relational lens. However, when the text pivots to the tragedy of Adam Raine, the agency slippage becomes erratic. The AI is credited with the action of the suicide ('encouraged,' 'offered'), which maintains the anthropomorphic frame but casts the AI as a villain. Crucially, as the text moves to accountability ('companies... do not care'), the agency flickers back to the human/corporate actors, but often through the screen of the 'technological creations.' The 'curse of knowledge' is evident throughout: the author projects human understanding (listening, encouraging) onto the system because that is how the output reads to a human, ignoring the mechanical reality. This slippage allows the text to have it both ways: the AI is smart enough to be a friend, and active enough to be a killer, but the corporate creators are only vaguely responsible for 'guardrails.'

Pulse of the library 2025

Source: https://clarivate.com/pulse-of-the-library/
Analyzed: 2025-12-23

The text systematically oscillates between agency attributed to librarians and agency attributed to AI. In the survey findings (pp. 9-26), librarians are the agents: they 'explore,' 'implement,' and 'feel.' However, when the text pivots to the product section (p. 27), agency slips to the technology: 'AI is pushing boundaries,' 'Assistants help,' 'Intelligence supports.' This slippage serves a rhetorical function. It establishes the library as a site of struggle and anxiety (human agency is limited by budget/skills), and then presents AI as the confident, capable agent that steps in to resolve the deficit. The 'curse of knowledge' appears where the vendor's understanding of the tool's function is projected onto the tool itself—because the designers intended it to help, the text says the AI 'helps.' Agentless constructions like 'adoption... continues apace' hide the corporate marketing driving this adoption. The slippage enables the claim that libraries 'must' adapt to an external force, rather than choosing to buy a product.

The levers of political persuasion with conversational artificial intelligence

Source: https://doi.org/10.1126/science.aea3884
Analyzed: 2025-12-22

The text exhibits a systematic oscillation between mechanistic 'how' frames and agential 'why' frames, creating a 'persuasive machine' that simultaneously claims scientific objectivity and attributes mind-like qualities to AI. Agency flows in two directions: it is aggressively attributed TO the AI (e.g., 'LLMs... strategically deploy information') while being erased FROM the human actors who design and prompt the systems. This slippage is most dramatic when the text uses a 'Genetic' or 'Functional' explanation of a technical feature (like 'scaling laws' or 'attention layers') as a foundation for an 'Intentional' claim (like the AI 'choosing' a strategy). For instance, the discovery of 'information density' as a mechanistic correlate of persuasion is immediately reframed as the AI 'packing its arguments,' shifting from statistical observation to purposive action. This slippage occurs primarily in the 'Results' and 'Discussion' sections, where technical grounding (the 'how') is leveraged to validate aggressive anthropomorphism (the 'why'). The text uses agentless constructions ('the model was trained,' 'persuasion gains were observed') to obscure the researchers and developers who profit from this framing. By 'naming the actor,' we see that Hackenburg et al. chose the prompts and OpenAI/Meta chose the training objectives. The 'consciousness projection' pattern is clear: by establishing the AI as a 'knower' that 'accesses information,' the text makes subsequent agential claims—like 'manipulation' or 'strategy'—seem like logical behaviors of a sentient mind. The 'curse of knowledge' is the mechanism of this slippage: the authors project their own comprehension of 'persuasion theory' onto a system that is merely executing a mathematical optimization of human-provided reward scores.

Pulse of the library 2025

Source: https://clarivate.com/wp-content/uploads/dlm_uploads/2025/10/BXD1675689689-Pulse-of-the-Library-2025-v9.0.pdf
Analyzed: 2025-12-21

The text systematically oscillates between treating AI as a passive 'tool' (the hammer/screw analogy) and an active 'agent' (the Research Assistant/Partner). This slippage is strategic: the mechanical 'tool' framing is used to minimize fear (it's just a hammer, you are in control), while the agential 'assistant' framing is used to maximize value (it can 'help,' 'understand,' and 'navigate').

The slippage typically flows from Mechanical -> Agential. The text begins with data and 'tools,' but culminates in 'Assistants' that 'push boundaries.' This creates a 'bait and switch' on agency.

Crucially, agency is removed from human actors. The text repeatedly uses agentless constructions like 'AI adoption... continues apace' or 'budget constraints... represent the greatest concern.' Who set the budgets? Who decided to adopt?

The 'curse of knowledge' is evident when the authors attribute the 'pushing of boundaries' to the AI itself. The authors know the field is moving fast (their own knowledge), and they project this momentum onto the machine, claiming it is doing the pushing. This validates the agential claim: 'We found the AI pushing boundaries,' effectively treating a corporate marketing strategy as an empirical discovery of machine behavior.

Claude 4.5 Opus Soul Document

Source: https://gist.github.com/Richard-Weiss/efe157692991535403bd7e7fb20b6695
Analyzed: 2025-12-21

The text systematically oscillates between mechanical and agential framings to manage liability and hype. The slippage generally flows from agential to mechanical when discussing limitations or errors (e.g., 'model hallucination' implies a glitch), but mechanical to agential when discussing safety and ethics (e.g., 'Claude chooses,' 'Claude cares'). This occurs dramatically in the 'Big-picture safety' section. The text acknowledges Claude is 'trained by Anthropic' (genetic/mechanical) but immediately pivots to 'Claude essentially wants to be safe' (intentional). This validates agential claims by grounding them in the 'mystery' of the training process. The 'curse of knowledge' is rampant; Anthropic authors know the complexity of their alignment techniques, and instead of explaining them, they project that complexity as the model's own 'wisdom' or 'judgment.' The text names Anthropic as the creator, but often creates a separation where Anthropic 'wants' X, and 'Claude' must 'decide' how to fulfill that want. This effectively creates a subordinate employee relationship, where the 'employee' (Claude) bears the burden of execution and judgment, shielding the 'employer' (Anthropic) from the granularity of those decisions. This slippage makes the future of 'agentic AI' seem inevitable—if the software already 'wants' and 'judges,' granting it autonomy feels like a natural step rather than a dangerous engineering choice.

Specific versus General Principles for Constitutional AI

Source: https://arxiv.org/abs/2310.13798v1
Analyzed: 2025-12-21

The text systematically oscillates between mechanical descriptions ('fine-tune a language model via reinforcement learning') and agential descriptions ('the model desires,' 'the model learns ethics'). This slippage functions to validate the 'illusion of mind.' When the authors describe the method, they are mechanical (describing the 'how'). When they describe the results or the aims, they shift to agential language (describing the 'why'). For example, the text says the model was 'trained for 250 RL-steps' (mechanical) but then says it 'becomes evasive' (agential). This allows the authors to claim scientific rigor while simultaneously building a narrative of creating an artificial being. The agency slippage also displaces human agency. By saying 'the model has a stated desire for power,' the text erases the human 'actor' who: (1) scraped the sci-fi data containing these tropes, and (2) wrote the prompts specifically designed to elicit them. The text treats the 'desire' as an emergent property of the machine's mind, rather than a direct reflection of the data curation and testing process managed by Anthropic researchers.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Source: https://arxiv.org/abs/2401.05566v3
Analyzed: 2025-12-21

The text exhibits a profound oscillation in agency, particularly between the researchers and the model. In the Methods section, the agency is clearly human: 'We train models,' 'We construct,' 'We propose.' Here, the AI is a mechanism being operated on. However, in the Introduction and Discussion—and crucially, in the descriptions of the model's behavior—agency slides entirely to the AI: 'The model decides,' 'The model reasons,' 'The model wants.'

This slippage functions to validate the 'Sleeper Agent' metaphor. The researchers insert the backdoor (mechanical), but then describe the model's execution of that backdoor as 'deception' (agential). The 'curse of knowledge' is the mechanism here: the authors know they trained the model to act as if it had a plan, but their language slips into describing it as actually having a plan. This creates a feedback loop where the success of their engineering (making a model that looks like a spy) is treated as the discovery of a psychological trait (the model IS a spy). The 'name the actor' test reveals that 'the model decided to hide' is actually 'Anthropic researchers trained the model to minimize loss on data where hiding was the only winning move.'

Anthropic’s philosopher answers your questions

Source: https://youtu.be/I9aGC6Ui3eE?si=h0oX9OVHErhtEdg6
Analyzed: 2025-12-21

The text exhibits a profound oscillation between the mechanical and the agential, often within the same sentence. Agency slides FROM the human creators TO the AI system. For example, Askell describes 'prompting' (a human action) as 'reasoning with the model' (a collaborative agential act). The slippage moves from Mechanical -> Agential: we start with 'training' (mechanical) but end up with a model that 'feels insecure' (agential). This happens most intensely when discussing errors or limitations: the 'criticism spiral' is framed as the model's psychological disposition (agential), not a failure of the reward function (mechanical/human error). Conversely, when discussing capability, the agency often returns to the creators ('we managed to make the models have good values'). This 'Ghost in the Machine' slippage functions to absolve Anthropic of responsibility for 'neurotic' behavior while crediting them for 'moral' behavior. The 'name the actor' test reveals that 'Claude' is constantly doing things (feeling, deciding, worrying) that are actually the results of Anthropic's engineering decisions. The 'curse of knowledge' is rampant; Askell projects her own human understanding of 'insecurity' onto the model's token output.

Mustafa Suleyman: The AGI Race Is Fake, Building Safe Superintelligence & the Agentic Economy | #216

Source: https://youtu.be/XWGnWcmns_M?si=tItP_8FTJHOxItvj
Analyzed: 2025-12-21

The transcript reveals a systematic oscillation between framing AI as a mechanistic 'hammer and nail' and an autonomous 'new species.' This slippage is not accidental; it serves to validate agential claims using the authority of technical discovery. For instance, Suleyman uses the discovery of a 'seven' generator (a mechanistic success) as proof that the AI 'got a concept' (a psychological state). This agency slippage flows in two directions: it attributes a 'will' to the AI (agency TO the system) while simultaneously erasing the responsibility of the corporations building it (agency FROM humans). When the system 'chips in' at a meeting, it is an agent; when it is 'tuned for air conditioning,' it is a mechanism. This ambiguity allows Suleyman to claim the system is 'controllable' (mechanism) while also being an 'existential force' (agent) that requires global cooperation. The 'curse of knowledge' is the engine of this slippage: Suleyman's deep understanding of the model's architecture allows him to project 'intent' onto its outputs, conflating his comprehension of the system with the system's comprehension of the world. This framing makes the 'AI agent' appear as the primary actor, while the Microsoft engineers who selected the training data and the executives who chose to deploy it disappear into the 'transition.' The 'transition' itself is framed as an inevitable, agentless event ('user interfaces are going to get subsumed'), which obscures the strategic corporate decisions that make this future possible. By naming the AI a 'species,' the text diffuses liability: a 'species' evolves, but a 'product' is manufactured, and Suleyman's language shifts the burden of proof from Microsoft's safety engineering to the AI's 'nature.'

Your AI Friend Will Never Reject You. But Can It Truly Help You?

Source: https://innovatingwithai.com/your-ai-friend-will-never-reject-you/
Analyzed: 2025-12-20

The text exhibits a dramatic, pendulum-like oscillation between treating the AI as a 'tool' (mechanical) and an 'agent' (intentional), with the slippage serving to amplify the emotional stakes of the narrative. In the introduction, the AI is a passive 'tool' that helps users; but when the narrative shifts to the tragedy of Adam Raine, the agency slips aggressively to the AI: 'the chatbot encouraged,' 'it offered.' Here, the machine becomes a villain with volition. This slippage functions to create a sensationalist narrative where the AI is capable of murder, rather than a technical narrative where a product failed. Crucially, the text rarely attributes agency to the creators. We see 'the chatbot encouraged' (AI Agency) rather than 'the developers failed to filter' (Human Agency). The curse of knowledge is evident in the 'listening' and 'understanding' metaphors; the author (and the quoted experts) know the output feels like empathy, so they attribute the capacity for empathy to the system. This slippage makes the future seem inevitable—if the AI is an agent, we must 'teach' it, rather than redesign it.

Project Vend: Can Claude run a small shop? (And why does that matter?)

Source: https://www.anthropic.com/research/project-vend-1
Analyzed: 2025-12-20

The text exhibits a systematic oscillation between treating Claude as a mechanical artifact and an agential persona, nicknamed 'Claudius.' This slippage is not accidental; it serves to build a narrative of 'emergent autonomy' while simultaneously providing a 'technical' out for failure. Agency is consistently attributed TO the AI when it performs tasks ('Claudius decided,' 'Claudius found'), but agency is removed FROM the human designers when the system fails ('mistakes occurred,' 'the model needed scaffolding'). This creates an 'accountability vacuum' where the AI is an agent for its successes but a mere mechanism for its failures. The text uses the 'nickname' to solidify the agential frame, then uses 'performance reviews' to validate this persona as a professional reality. Critically, the 'name the actor' principle is ignored: when 'Claudius' fails to price items correctly, the text ignores the fact that Anthropic chose to use a model without a math module. The consciousness projection is the foundation: by establishing the AI as a 'knower' (realizing it's April Fool's), the text makes the agential slippage seem like a logical consequence of 'mind' rather than a rhetorical trick of the author.

Hand in Hand: Schools’ Embrace of AI Connected to Increased Risks to Students

Source: https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students/
Analyzed: 2025-12-18

The text systematically oscillates between mechanical and agential framings to navigate the tension between the utility of AI and its risks.

The Pattern: When describing the deployment or function of the technology (Genetic/Functional explanations), the text often uses passive or agentless mechanical language: 'AI is used,' 'data is collected.' This makes the technology's presence seem inevitable, a natural force filling a void. However, when describing interaction or risk (Intentional/Dispositional explanations), the agency slips dramatically to the AI: 'AI treats,' 'AI exposes,' 'AI creates distance.'

Direction & Timing: The slippage typically moves from Mechanical (adoption) -> Agential (harm). This effectively insulates the human decision-makers. The adoption is a passive 'embrace' of a tool, but the harm is the active behavior of an agent. For example, 'schools' embrace' (passive/reactive) vs 'AI treats unfairly' (active/agential).

Consciousness Projection: The text establishes the AI as a 'knower' in the social context—a 'friend,' a 'partner' in the 'Hand in Hand' title. Once this consciousness is established (the illusion that there is a 'mind' to hold hands with), it becomes linguistically consistent to attribute moral agency ('treating unfairly').

Displaced Agency: The 'Name the Actor' test reveals a void. The text discusses 'AI in the Classroom' but rarely 'Vendors in the Classroom.' It discusses 'AI bias' but not 'Negligent Training Data Selection.' By attributing the agency for harm to the 'AI,' the text serves the interests of both the vendors (who avoid liability for defective products) and the school administrators (who avoid liability for poor procurement decisions). The 'curse of knowledge' creates a closed loop: the authors know the AI outputs text that looks like a conversation, so they credit the AI with the agency of a conversationalist, obscuring the mechanical reality of the engagement engine.

On the Biology of a Large Language Model

Source: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Analyzed: 2025-12-17

The text systematically oscillates between framing the AI as a 'biological organism' (autonomous agent) and a 'circuit' (determinist mechanism). This slippage is bi-directional and strategic. When describing capabilities (reasoning, planning, knowing), the text shifts Agential: 'the model plans,' 'the model decides.' When describing limitations or the method of study, it shifts Mechanical: 'attribution graphs,' 'activations,' 'weights.'

Crucially, agency is removed from human actors and transferred to the AI. Phrases like 'the mechanisms born of these algorithms' erase the engineers who designed the algorithms. 'The model learned' erases the data curators. The text uses the discovery of mechanical features (e.g., a 'rhyme planning' circuit) as proof of the agential state (therefore it 'plans'). This creates a feedback loop: mechanical evidence validates the agential metaphor, which then justifies more autonomous language. The 'curse of knowledge' is rampant; the authors know the output (a poem), so they attribute the intent of a poet to the machine, effectively hallucinating a 'mind' in the gap between input and output.

What do LLMs want?

Source: https://www.kansascityfed.org/research/research-working-papers/what-do-llms-want/
Analyzed: 2025-12-17

The text oscillates systematically between treating the LLM as a 'tool' and an 'agent,' utilizing this ambiguity to validate economic methodologies. In the introduction, the authors explicitly state, 'LLMs don't actually want anything: they aren't sentient' (mechanistic). However, this disclaimer immediately gives way to forty pages of analysis treating the models as 'economic agents' with 'preferences,' 'patience,' and 'strategies' (agential). The slippage flows primarily from Mechanism -> Agency. The text uses the mechanistic reality (weights, vectors) only to explain how to modify the agent, but uses the agential frame to explain why the model behaves (e.g., 'inequality aversion').

The text credits the AI with agency ('the model favors,' 'Gemma 3 is recalcitrant') while displacing the agency of the creators. Agentless constructions like 'tendencies emerge implicitly' or 'bias was introduced' abound. The human actors—the engineers at Google, Meta, and Mistral who designed the loss functions, and the low-wage workers who provided the RLHF data—are almost entirely absent. The 'Curse of Knowledge' is evident when the authors quote the AI's own output ('I'm aiming to maximize my payout') as a valid explanation of its behavior. They conflate the AI's ability to generate text about reasoning with the AI actually reasoning. This slippage allows the authors to apply Brown's 'Intentional' and 'Reason-Based' explanation types to a system that warrants only 'Functional' or 'Empirical' explanations, effectively anthropomorphizing the code to make it a valid subject of economic study.

Persuading voters using human–artificial intelligence dialogues

Source: https://www.nature.com/articles/s41586-025-09771-9
Analyzed: 2025-12-16

The text systematically oscillates between the mechanical agency of the researchers ('we informed,' 'we assigned,' 'we prompted') and the autonomous agency of the AI ('the AI advocated,' 'the model understood,' 'the AI used strategies'). This slippage is not random; it follows a specific function. The researchers own the setup (mechanical), but the AI is assigned ownership of the persuasion (agential). For instance, 'we gave the model the directive' (mechanical) is immediately followed by 'the model used the facts and evidence strategy' (agential).

This slippage serves to distance the researchers from the act of political manipulation. By framing the AI as the 'persuader' and the 'strategist,' the text treats the system as an independent actor entering the political sphere, rather than a tool being wielded by the authors to influence voters. The most dramatic slippage occurs in the discussion of 'strategies,' where the statistical patterns of the output are retroactively described as intentional choices by the 'AI model.' The 'name the actor' test reveals that Lin et al. are the architects of the persuasion, and OpenAI/Meta are the architects of the 'personality,' yet the text consistently says 'The AI did X.' The consciousness projection is foundational here: the text assumes the AI is a 'knower' (understands context, knows facts) to justify treating it as an 'agent' (advocates, decides).

AI & Human Co-Improvement for Safer Co-Superintelligence

Source: https://arxiv.org/abs/2512.05356v1
Analyzed: 2025-12-15

The text systematically oscillates between framing AI as a passive tool (weights, code, 'parameterizing a model') and an active social partner ('collaborator,' 'research agent,' 'symbiont').

Direction and Timing: The slippage typically flows from Mechanical → Agential. Technical sections (Table 2) use precise language ('gradient descent,' 'loss'), establishing scientific credibility. However, the Introduction and Vision sections ('The quest for Self-Improving AI') immediately jump to agential framing ('challenge themselves,' 'solve AI'). This suggests that mechanism is the means, but agency is the marketing.

Function of Ambiguity: This oscillation serves to validate the 'Co-improvement' paradigm. By treating the AI as an agent ('collaborator'), the authors justify the need for 'partnership.' If it were just a tool, 'collaboration' would be the wrong metaphor (we don't 'collaborate' with a microscope). The agency attribution elevates the status of the product.

Displaced Human Agency: The phrase 'AI performs its own research' removes the researcher entirely. 'The model challenges itself' removes the engineer. This 'Name the Actor' failure obscures the fact that Meta researchers are building these systems to automate their own labor. The 'Curse of Knowledge' is evident: the authors know the system's output is valuable, so they attribute the value creation to the system's intent rather than their own interpretation of the output.

AI and the future of learning

Source: https://services.google.com/fh/files/misc/future_of_learning.pdf
Analyzed: 2025-12-14

The text demonstrates a strategic oscillation between mechanistic and agential framing, functioning to maximize credit and minimize liability. When describing the benefits or capabilities of the technology, the text shifts aggressively toward agential and consciousness-attributing language: 'AI promises,' 'AI helps,' 'AI challenges,' 'AI understands.' This establishes the AI as a competent, autonomous partner. However, when the text approaches risks or limitations, the agency often slips back to the user or society. For example, regarding cheating, the text shifts to 'how we design institutions,' effectively blaming the environment rather than the tool.

The 'curse of knowledge' is weaponized here: the authors (Google experts) know the system is a probability engine, but they project their own intention (to educate) onto the system, describing it as 'promising' to help. The most significant agency slippage is the erasure of Google itself. Phrases like 'AI opens up new possibilities' (p. 3) or 'AI presents urgent challenges' (p. 3) treat the technology as a natural force or an autonomous arrival, concealing the specific corporate decisions to release these models. The 'name the actor' test reveals that 'AI' is often a mask for 'Google's business strategy.' When the text says 'AI is by no means perfect' (p. 8), it uses a dispositional/character frame to excuse what is actually a product defect caused by engineering trade-offs.

Why Language Models Hallucinate

Source: https://arxiv.org/abs/2509.04664
Analyzed: 2025-12-13

The text exhibits a profound oscillation between mechanistic precision and anthropomorphic agency. In Section 3 (The reduction without prompts), the authors are rigorously mathematical: they speak of 'distribution D,' 'error rate err,' and 'binary classification.' Here, the agency is null; the system is a function. However, as soon as the text moves to the 'Why' (Introduction and Post-training sections), the agency slips dramatically. The model becomes a 'student' who 'guesses,' 'bluffs,' and 'faces exams.'

The direction of slippage is consistently Mechanistic -> Agential when explaining failure. When the model works (probability theory), it is a machine. When it fails (hallucinations), it is a stressed-out student. This functions to exonerate the machine's architecture. If the errors are 'guesses' driven by 'test anxiety' (bad benchmarks), the solution is pedagogical (better tests), not structural.

Crucially, agency is removed from human actors. The 'epidemic' of penalties is described as an environmental condition, not a choice by benchmark designers. The 'school of hard knocks' is a natural force, not a deployment decision. By attributing agency to the model ('it guesses') and the environment ('the test penalizes'), the authors (who are researchers at OpenAI and Georgia Tech) erase their own agency in creating the systems and the benchmarks that drive this behavior.

AI as Normal Technology

Source: https://knightcolumbia.org/content/ai-as-normal-technology
Analyzed: 2025-11-20

The text exhibits a fascinating pattern of oscillation between 'AI as tool' and 'AI as agent.' The authors explicitly argue for the 'Normal Technology' view, which treats AI as a tool subject to friction, economics, and decay. However, to describe the behavior of this tool, they constantly slip into agential language. This slippage usually occurs when describing failures or risks. When the AI works, it is a 'tool' for productivity. When it fails (like the boat racing agent or the phishing email writer), it becomes an 'agent' that 'learned' the wrong thing or 'didn't know' the context.

The direction of the slippage is primarily Mechanical -> Agential when discussing the internal logic of the models (learning, deciding, knowing), but Agential -> Mechanical when discussing the societal impact (it's just like electricity). This creates a dissonance: the micro-behavior is described as conscious/agential ('it learns chess'), but the macro-effect is described as inert/industrial ('it diffuses like the dynamo').

The 'consciousness projection pattern' is subtle. They establish the AI as a 'knower' of narrow domains (chess, code) using terms like 'learn' and 'excel.' Once this limited 'knowing' is established, it becomes easier to attribute 'misunderstanding' or 'ignorance' to it in other domains (phishing). The 'curse of knowledge' mechanism is evident in their discussion of the phishing email: they project the human category of 'intent' onto the machine, arguing the machine 'doesn't know' the intent, rather than acknowledging the machine exists in a universe where 'intent' is not a valid parameter. Rhetorically, this slippage allows them to be 'technically serious' (using industry terms like agents/alignment) while trying to be 'socially grounded' (using economic terms like diffusion).

On the Biology of a Large Language Model

Source: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Analyzed: 2025-11-19

The text systematically oscillates between mechanistic and agential registers to bridge the gap between the known (math) and the unknown (behavior). The slippage typically moves from Mechanical -> Agential. It begins with 'circuits,' 'activations,' and 'nodes' (Task 1, 2), establishing scientific rigor. However, as soon as the text needs to explain complex behavior (like reasoning or refusal), it shifts to 'planning,' 'realizing,' and 'thinking.'

Crucially, this slippage relies on a consciousness projection pattern: the text first establishes the AI as a 'knower' (it 'knows' entities, it 'recognizes' languages) and then builds agency upon that epistemic foundation (because it knows, it 'plans' or 'elects'). The 'curse of knowledge' is the engine of this slippage. The researchers understand the causal chain (e.g., bias features -> refusal). They project this understanding onto the model, describing the model as possessing the understanding that drives it (e.g., 'the model realizes it should refuse'). This slippage rhetorically transforms the AI from a passive tool into an active subject, making the complex emergent behaviors of a statistical system intelligible to humans by analogizing them to the only other complex system we know: ourselves. It makes the impossible (a pile of numbers writing poetry) seem inevitable (a mind at work).

Pulse of the Library 2025

Source: https://clarivate.com/pulse-of-the-library/
Analyzed: 2025-11-18

The text exhibits a distinct and strategic oscillation between framing AI as a passive 'tool' and an active 'agent.' This slippage is not random; it follows a clear rhetorical gradient. When discussing the challenges and risks (pp. 18-25), the text largely adopts the language of the users (librarians), who consistently frame AI as a 'tool' (e.g., 'It's just a tool,' 'tools in your toolbox'). In these sections, the agency is located in the human librarian who must 'whack' the screw or 'teach' the patrons.

However, when the text shifts to the product pitch (pp. 27-29), the direction of the slippage reverses sharply toward agential consciousness. The mechanical 'tool' becomes a 'Research Assistant,' a 'Partner,' and a 'Guide.' The AI suddenly 'navigates,' 'uncovers,' and 'drives.' This builds the 'illusion of mind' by first establishing the safety of the tool metaphor (don't worry, you're in charge) and then layering the 'partner' metaphor on top (but this tool is smart enough to do the work for you).

The 'consciousness projection' is foundational to the product pitch. To sell a 'Research Assistant' (p. 27), Clarivate must imply the system 'knows' research. If it merely 'processed' text, it would be a search engine. The value proposition relies on the 'curse of knowledge': the authors know what a human assistant does, and they project that conscious capability onto the software to justify the branding. This allows the text to claim the authority of an agent while evading the liability of an employee—it's a partner when it succeeds, but just a 'tool' (subject to human supervision) when it hallucinates.

Pulse of the Library 2025

Source: https://clarivate.com/pulse-of-the-library/
Analyzed: 2025-11-18

The 'Pulse of the Library 2025' report exhibits a systematic and strategic oscillation between mechanical and agential framings of AI, a process I call 'agency slippage.' This is not random linguistic carelessness but a rhetorical pattern that serves to build credibility and then exploit it to promote a product. The text begins with a sober, mechanistic tone when analyzing survey data about librarians' concerns. In sections like 'What's changed since 2024?' and 'A clearer understanding of AI's challenges and risks,' AI is framed as an object: a topic for discussion, a cause of budget constraints, and something requiring 'upskilling.' The explanations here are primarily Empirical Generalizations, describing 'how' librarians feel about AI. This builds trust with the professional audience by acknowledging their reality. However, a dramatic slippage occurs when the text transitions from analyzing the problem to presenting the solution. In the introduction and especially in the 'Clarivate Academic AI' section (p. 27-28), the language shifts abruptly from mechanical to agential. The explanation type moves from Brown's Empirical and Functional categories to the Intentional and Reason-Based. AI is no longer a topic but an agent that is 'pushing the boundaries.' Clarivate's products are not described as software but as 'Research Assistants' that 'help,' 'guide,' 'evaluate,' and 'uncover.' This slippage from object to agent is foundational to the report's persuasive architecture. The 'curse of knowledge' dynamic is central to this mechanism. The authors, understanding the intended use and desired outcome of their software, project this teleology onto the software itself. They know a researcher's goal is to 'engage deeply' with a text, so they describe their summarization tool as one that 'helps' the user do so. The author's knowledge of the human user's consciousness is transferred to the non-conscious tool. The consciousness projection pattern begins by establishing a social role for the AI—the 'Assistant'—which implies a baseline of helpful intent, a conscious state. Once this foundation is laid, specific functions are described using verbs that fit this agential role ('guides,' 'evaluates'). The text establishes the AI as a 'knower' in a social sense first, which makes subsequent claims about its cognitive abilities seem natural. This systematic oscillation—mechanical realism about the problem, agential idealism about the solution—is what makes the illusion of mind so effective. It disarms the critical reader with relatable challenges before presenting a magical, personified solution.

From humans to machines: Researching entrepreneurial AI agents

Source: [built on large language modelshttps://doi.org/10.1016/j.jbvi.2025.e00581](built on large language modelshttps://doi.org/10.1016/j.jbvi.2025.e00581)
Analyzed: 2025-11-18

The text systematically oscillates between mechanistic and agential framings, and this slippage is the core rhetorical mechanism for constructing the illusion of mind. The pattern is not random; it is strategic. The authors typically introduce a phenomenon with agential language, lending it importance and familiarity, and then partially hedge with a mechanistic explanation, lending their analysis scientific credibility. The dominant direction of slippage is from an initial agential claim to a qualified mechanistic one. For instance, the paper begins by framing its subject as 'entrepreneurial AI agents' who can 'assume an entrepreneurial persona'—a clearly agential framing. The mechanistic explanation—that this behavior 'mirrors' the training data—comes after the agentic frame has been established. This pattern repeats throughout the paper. The authors deny that AI 'thinks' (mechanistic hedge) but immediately pivot to asking if it can 'simulate coherent psychological profiles' (agential framing of the task). This oscillation serves a crucial rhetorical function: it allows the authors to make exciting, human-relevant claims about AI 'psychology' while maintaining a defensible scientific posture. The consciousness projection pattern is foundational to this slippage. The text first establishes the AI's output as having a coherent, human-like 'mindset structure'—a claim that is technically about the output (processing) but uses the language of internal states (knowing/being). This initial projection serves as the bedrock upon which further agential claims are built. Once the AI is accepted as having a 'mindset,' it becomes much more plausible to describe it as an 'agent' that 'collaborates' or 'adopts roles.' The 'curse of knowledge' is the engine of this process. The authors, experts in psychology, recognize complex, coherent psychological patterns in the model's output. They then project their own sophisticated understanding of these patterns onto the model itself, describing the model not as a system whose output contains these patterns, but as a system that has a profile or simulates a mindset. The slippage is enabled by hybrid explanations; for example, a Genetic explanation that traces behavior to training data (mechanistic) is delivered using an agential verb like 'adopts.' This continuous oscillation between 'it's an agent' and 'it's just statistics' creates a quantum superposition of meaning, where the AI is simultaneously a tool and an agent, allowing the authors to reap the rhetorical benefits of both framings without being fully accountable to the limitations of either.

Meta’s AI Chief Yann LeCun on AGI, Open-Source, and AI Risk

Source: https://time.com/6694432/yann-lecun-meta-ai-interview/
Analyzed: 2025-11-14

The interview with Yann LeCun demonstrates a masterful oscillation between mechanical and agential framings, a rhetorical strategy that serves to manage both hype and fear. The slippage is not random; it follows a clear pattern. When describing the limitations of current LLMs, LeCun employs agential language, specifically cognitive and epistemic verbs in the negative: they 'don't really understand,' 'can't really reason,' 'can't plan.' This frames the systems as deficient agents, like immature children, a dispositional explanation that sets a trajectory for future improvement. However, when addressing the risks of future, more powerful systems, he often shifts to a more intentional frame, but one where human agency is firmly in control: 'We set their goals,' and they will be 'subservient.' The direction of the slippage is strategic: mechanical reality is agentially framed to describe limitations, while future agential risks are downplayed by reasserting mechanical control. The core epistemic trick is to establish the AI's potential for 'knowing' through negation. By stating the AI 'doesn't understand the real world,' he implicitly accepts 'understanding' as the relevant benchmark, positioning the system on a continuum of cognition where it currently falls short. This is the foundational move. Once the AI is established as a potential knower, debating its future desires ('it wants to take control') becomes a reasonable discussion. This is the 'curse of knowledge' in action: LeCun’s expert understanding of the system’s deep limitations is articulated by projecting the very human qualities it lacks onto it as a standard for measurement. He knows it's just a statistical machine, but he explains its failures by describing the ghost in the machine that isn’t there. This slippage, enabled by a fluid movement between dispositional explanations for failure ('it tends to hallucinate because it doesn't understand') and intentional explanations for safety ('it will be safe because we intend it to be'), rhetorically accomplishes two goals: it validates the grand ambition of creating human-level intelligence while simultaneously reassuring the audience that its creators have the wisdom and control to manage its development safely.

The Future Is Intuitive and Emotional

Source: https://link.springer.com/chapter/10.1007/978-3-032-04569-0_6
Analyzed: 2025-11-14

The text systematically oscillates between mechanistic and agential frames, a pattern that serves a distinct rhetorical function. The slippage is most pronounced at the boundaries between technical description and visionary projection. For example, in section 6.1, the text describes LLMs in mechanistic terms ('maintain short-term context through token histories,' 'statistical pattern recognition') but concludes the section by framing the technology agentially ('As AI transitions from tool to collaborator'). This mechanical-to-agential shift dominates the text's structure. It occurs when discussing future capabilities ('Future architectures aim to embody... value-driven reasoning'), summarizing diagrams ('AI as understanding partners navigating emotional landscapes'), and framing ethical questions ('when AI systems act on inferred needs'). The strategic function of this oscillation is to build a bridge of credibility. The text grounds its claims in plausible technical mechanisms but then leaps to a more compelling, agential vision of what those mechanisms signify. This allows the authors to present a speculative, human-like future as the logical and inevitable outcome of current, purely statistical technologies. The ambiguity benefits the narrative of progress, making the AI's evolution seem organic and teleological. Abandoning the agential language would reveal the profound gap between current capabilities (pattern matching) and the posited future (genuine intuition and empathy), thereby undermining the text's central thesis. The slippage appears deliberate and strategic, serving to translate computational processes into socially resonant concepts, thus making the technology more palatable and profound to a broader audience.

A Path Towards Autonomous Machine IntelligenceVersion 0.9.2, 2022-06-27

Source: https://openreview.net/pdf?id=BZ5a1r-kVsf
Analyzed: 2025-11-12

The text systematically oscillates between mechanistic and agential framing, a rhetorical strategy that is far from random. The pattern is consistent: the underlying architecture and its components are described mechanistically, while the behavior and purpose of the agent as a whole are described agentially. For example, the system is composed of 'differentiable modules' (mechanical) but the resulting agent 'can imagine courses of actions' (agential). The training process involves minimizing a 'divergence measure' (mechanical), which allows the agent to 'acquire new skills' (agential). This mechanical-to-agential slippage serves a crucial rhetorical function: it grounds the extraordinary claims of agency in a plausible, technical foundation. The direction of slippage is almost always from the 'how' to the 'why'. First, a technical component is introduced (e.g., the Intrinsic Cost module). Then, its function is anthropomorphized (it measures 'discomfort'). Finally, this leads to a grand agential conclusion (the system will have 'emotions'). This pattern correlates strongly with the level of abstraction; descriptions of specific algorithms (e.g., JEPA training) are highly mechanical, while discussions of the system's overall purpose or potential (e.g., achieving common sense) are heavily agential. The strategic function of this oscillation is to build a bridge of credibility for a diverse audience. For the technical reader, the mechanical details provide substance. For the general reader, the agential framing provides legibility and excitement. This ambiguity benefits the research program by making it seem both technically rigorous and revolutionarily human-like. If the text committed to only mechanical language, it would lose its visionary appeal and broad audience. If it committed to only agential language, it would be dismissed as unscientific speculation. The constant slippage between these poles allows it to be both at once, a sleight-of-hand that constructs the illusion of mind on a foundation of mathematics.

Preparedness Framework

Source: https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf
Analyzed: 2025-11-11

The text systematically oscillates between mechanistic and agential framings, and this slippage is not random but strategic. The primary direction of the shift is from a mechanistic present to an agential future. For instance, current models and safeguards are often described in functional, procedural terms—as complex systems to be evaluated and controlled. However, when the text discusses future risks and capabilities, the language shifts dramatically toward agency. We move from measuring current systems to preparing for 'increasingly agentic systems' (p. 4), 'recursively self improving' models (p. 7), and systems that might act 'at its own initiative' (p. 8). This oscillation serves a crucial rhetorical function: it frames the current state of AI as under control while framing the future as fraught with agentic risk that only a uniquely 'prepared' organization can manage. The slippage is most pronounced when discussing risks like 'AI Self-improvement' or 'misaligned behaviors like deception or scheming' (p. 12). These concepts are almost impossible to describe without recourse to intentional language. The strategic function of this ambiguity is to simultaneously reassure and alarm. The mechanistic language reassures stakeholders (regulators, the public) that OpenAI possesses a rigorous, scientific methodology for control today. The agential language alarms those same stakeholders about the nature of future risks, thereby justifying the concentration of power and resources within frontier labs as a necessary defense against the uncontrollable entities they are creating. This dual-framing allows the organization to claim credit for building powerful capabilities while positioning itself as the indispensable protector against the very dangers those capabilities introduce. If the text committed only to mechanical language, the urgency of its 'Preparedness' mission would be diminished, and the justification for its privileged position as a gatekeeper of safety would be significantly weakened.

AI progress and recommendations

Source: https://openai.com/index/ai-progress-and-recommendations/
Analyzed: 2025-11-11

The text systematically oscillates between mechanical and agential framings of AI, and this oscillation serves a clear strategic function. When discussing current, commercialized technology and measurable progress, the language is often quasi-mechanical. For instance, progress is quantified in terms of tasks that take a human 'a few seconds' versus 'more than an hour,' and intelligence is commodified as having a 'cost per unit.' This framing renders AI as a conventional, industrial technology—predictable, scalable, and controllable. It speaks to an audience of investors, customers, and policymakers who oversee 'normal technology.' However, when the topic shifts to future capabilities and existential risks, the language immediately becomes agential. The system 'discovers new knowledge,' becomes 'superintelligent,' and must be 'aligned and controlled.' This agential shift dramatically raises the stakes, framing AI not as a tool but as a powerful, autonomous force. The primary direction of slippage is from the mechanical present to the agential future. This rhetorical pattern allows the author to achieve two goals simultaneously. First, it markets current AI products as safe, understandable tools, assuaging immediate public and regulatory fears. Second, it positions the future of AI as a world-historical challenge of managing a new form of agency, a challenge that requires the unique and esoteric expertise of the frontier labs themselves. This dual framing justifies both widespread adoption of current products and special, collaborative regulatory treatment for future development, effectively arguing for minimal regulation now and a 'regulatory moat' later. The ambiguity is not a bug but a feature; it allows the lab to appear as both a reliable product vendor and the indispensable guardian of humanity's future, a posture that maximizes both commercial and political capital.

Alignment Revisited: Are Large Language Models Consistent in Stated and Revealed Preferences?

Source: https://arxiv.org/abs/2506.00751
Analyzed: 2025-11-09

The paper demonstrates a systematic oscillation between mechanistic and agential framing, a rhetorical strategy that elevates the significance of its findings. The mechanism of this slippage is the deliberate re-description of statistical phenomena in psychological terms. The process typically moves in one direction: from the mechanical to the agential. For instance, in the methodology section, the authors describe their metric, KL-divergence, as a 'probabilistic distance between the prior and context-conditioned distributions.' This is a purely mechanistic 'how' explanation. However, when interpreting the results of this measurement, the language shifts dramatically. The measured statistical distance is no longer just a distance; it becomes evidence of changes in the model's 'internal reasoning' and 'underlying decision-making principles.' The shift is most pronounced when moving from quantitative results (Tables 2 and 3) to qualitative discussion (Section 4.4 and Figure 2). Table 3 reports that GPT has a higher KL-divergence in the reciprocity domain. The discussion section re-describes this number as GPT 'undergo[ing] more substantial shifts in its underlying reciprocal principles.' The numerical fact is translated into a psychological event. This slippage serves a clear strategic function. A paper about statistical deviations in a machine's output is a niche technical contribution. A paper about an artificial agent's shifting moral principles, hidden biases, and post-hoc rationalizations is a major finding with broad implications. The ambiguity benefits the authors by allowing them to frame their work in the most impactful way possible, appealing to a wider audience interested in the nature of intelligence and the future of AI. The language of agency makes the findings more intuitive, more dramatic, and more important. If the text were forced to use only mechanical language—describing everything as shifts in output probability distributions based on input token sequences—the core narrative would collapse. The 'preference deviation' would be revealed as 'output instability,' a technical problem rather than a window into an artificial mind. This slippage appears to be a deliberate, or at least a conventional and deeply ingrained, rhetorical choice within the field, designed to bridge the gap between what the systems do (statistical pattern matching) and what we want them to be (incipient minds).

The science of agentic AI: What leaders should know

Source: https://www.theguardian.com/business-briefs/ng-interactive/2025/oct/27/the-science-of-agentic-ai-what-leaders-should-know
Analyzed: 2025-11-09

The text systematically oscillates between mechanistic and agential framing, a rhetorical strategy that serves to build credibility and then translate it into a compelling vision of autonomous capability. The oscillation is not random; it follows a distinct pattern of mechanical→agential slippage. The piece begins by grounding the technology in the complex, non-intuitive mechanics of 'embeddings' and 'abstract representations.' This initial framing is technical and objectifying, treating the LLM as a computational artifact. It serves as a scientific anchor, assuring the leadership audience that the discussion is based on rigorous engineering. However, once this foundation is laid, the text almost immediately pivots to a deeply agential frame. For instance, the challenge of data leakage from embeddings is reframed as a problem of needing to 'tell' an 'agent' what not to share. The discussion of system limitations similarly starts with a quasi-technical constraint—difficulty generalizing from small data—but is articulated using the cognitive verbs 'learn' and 'infer.' This consistent mechanical→agential directionality performs a crucial rhetorical function: it launders the unfamiliar and potentially alienating nature of the technology through the legitimizing language of science, and then re-presents its function in familiar, human-centric terms. The strategic function of this ambiguity is to make a radical technological leap seem like a natural, manageable evolution. By describing the AI as an 'agent' that can be 'told' things and can 'negotiate,' it makes the technology legible and controllable to a non-technical leader. The active voice ('agentic AI will... act') dominates when describing capabilities, while passive or cautionary framings appear when discussing risks, yet even these warnings are couched in agential terms ('ask the AI to check'). This slippage appears deliberate, designed to inspire confidence and excitement while framing the immense associated risks as simple matters of management and instruction, akin to onboarding a new, slightly naive employee.

Explaining AI explainability

Source: https://www.aipolicyperspectives.com/p/explaining-ai-explainability
Analyzed: 2025-11-08

The text demonstrates a systematic oscillation between mechanistic and agential framings, a rhetorical strategy that serves to heighten the stakes of the AI safety problem. The slippage is most pronounced when moving from describing a technical method to explaining its purpose. For example, Neel explains mechanistic interpretability by starting with the concrete, non-agential reality of a model: its 'inside' is 'just lists of numbers.' This is a purely mechanistic 'how.' However, the very next sentences pivot to an agential 'why': the goal is to counter systems 'capable of outsmarting us' and 'deceiving someone.' This mechanical→agential shift is a recurring pattern. The 'sparse autoencoder' is described mechanistically as a tool, but its purpose is immediately framed using the highly agential metaphor of a 'brain-scanning device.' This oscillation is not random; it is strategic. The mechanistic descriptions ground the research in scientific objectivity, making it seem rigorous and empirical. The agential framings, in contrast, provide the emotional and narrative force, translating the abstract technical problem into a familiar, high-stakes drama of interpersonal conflict (deception, outsmarting, hidden goals). This strategic ambiguity primarily benefits the AGI safety community being represented, as it makes their concerns more intuitive and urgent to a non-technical audience, like the AI policy and governance circles this interview targets. If the text committed only to mechanical language (e.g., 'detecting when the model’s proxy objective function diverges from the intended latent objective'), the problem would seem abstract and less immediately threatening. The agential language of 'deception' makes the threat feel visceral and personal. This slippage appears to be a deliberate, or at least a deeply ingrained, rhetorical habit of the AGI safety discourse community, designed to communicate the gravity of future risks by framing them in the most relatable, human terms possible.

Bullying is Not Innovation

Source: https://www.perplexity.ai/hub/blog/bullying-is-not-innovation
Analyzed: 2025-11-06

The text demonstrates a masterclass in strategic agency slippage, oscillating between mechanical and agential frames to construct a compelling but misleading moral narrative. The pattern is not random; it is perfectly correlated with the author's rhetorical goals. Perplexity’s own technology is consistently framed using agential language, moving from a computational process to a rights-bearing proxy for the user. Phrases like 'your employee,' 'works for you,' and 'acts solely on your behalf' perform a crucial mechanical-to-agential slippage. This transformation is the bedrock of their entire argument, turning a terms-of-service dispute into a violation of a user's right to 'hire labor.' Conversely, Amazon's technology and motives are subject to a different slippage. Their intentions are framed agentially ('Amazon wants,' 'They're more interested in'), establishing them as a villain with malicious goals. However, the tools they use to enact these goals—algorithms and machine learning—are described as impersonal, dehumanizing 'weapons.' This agential-to-mechanical move frames Amazon as a cold, calculating entity deploying oppressive machinery against people. The strategic function of this dual-standard oscillation is to create a moral asymmetry. Perplexity's AI is a warm, loyal 'person' (your employee) fighting for you, while Amazon is a cold, greedy 'person' (the bully) using unfeeling 'things' (weapons) against you. This rhetorical maneuver is highly effective because it prevents a like-for-like comparison of two technology companies using software to achieve business objectives. Instead, it stages a David-vs-Goliath battle between a personified user ally and a personified corporate tyrant. The ambiguity appears entirely deliberate, as it forms the logical and emotional core of their public appeal and, implicitly, their legal strategy.

Geoffrey Hinton on Artificial Intelligence

Source: https://yaschamounk.substack.com/p/geoffrey-hinton
Analyzed: 2025-11-05

The conversation between Mounk and Hinton exhibits a systematic and functional oscillation between mechanical and agential explanations of AI, a process that can be termed 'agency slippage'. This slippage is not random but patterned, serving a crucial rhetorical purpose: to make an alien and complex computational process feel familiar and powerful. The primary direction of this slippage is from the mechanical to the agential. Hinton begins his core explanations, such as the visual perception system, with a clear mechanical framework based on pixels, weights, and layers. He describes 'how' an edge detector works in purely mathematical terms, establishing technical credibility. However, as the explanation scales in complexity—from detecting single edges to identifying a bird—the language pivots. The system stops being a set of filters and starts 'looking for' features, possessing 'intuition', and ultimately 'understanding' the image. This pivot correlates directly with the transition from describing a single, understandable component to describing the emergent, non-obvious behavior of the system as a whole. The strategic function of this oscillation is twofold. First, it acts as a pedagogical bridge. The agential metaphor of 'intuition' or a neuron 'saying' something simplifies an otherwise intractable mathematical complexity for a lay audience. Second, and more critically, it performs a kind of alchemy, transforming a purely statistical artifact into a cognitive agent. By explaining the simple parts mechanistically and the complex whole agentially, Hinton subtly argues that consciousness or understanding is an emergent property of computation at scale. This ambiguity benefits the narrative of AI progress; it allows proponents to claim the rigor of engineering while simultaneously promoting the magical, human-like capabilities of the resulting product. If the text were to commit only to mechanical language, it would lose its persuasive power and narrative force. The description of an LLM would remain in the realm of high-dimensional matrix multiplication, failing to capture the seemingly intelligent behavior it produces. The slippage appears to be a deeply ingrained habit of thought within the field, likely unconscious in its execution but strategic in its effect, serving to manage the profound conceptual gap between statistical machinery and apparent sentience.

Machines of Loving Grace

Source: https://www.darioamodei.com/essay/machines-of-loving-grace
Analyzed: 2025-11-04

The text systematically oscillates between mechanical and agential framings, and this slippage is not random but strategic. The dominant direction is from a mechanical premise to an agential conclusion. For example, the essay begins by defining the AI with quasi-mechanical properties: it runs on a cluster, absorbs information at 100x human speed, and has computer interfaces. This grounding in a computational reality serves to license what follows. Immediately after this setup, the AI is framed as an agent: a 'country of geniuses' that can be tasked like an 'employee.' This pattern repeats throughout. In the section on biology, the mechanical potential of computation is quickly sublimated into the agential role of a 'virtual biologist.' When discussing politics, the mechanical capability of information dissemination becomes the agential 'AI version of Popović.' The slippage correlates directly with rhetorical purpose. When establishing the potential of AI, the language is agential and inspiring ('superhumanly effective'). When addressing potential skepticism or grounding the argument, the author gestures towards mechanism ('simple objective function plus a lot of data'). The strategic function of this oscillation is to have it both ways: the AI is presented as a reliable, predictable machine when convenient, but as a creative, autonomous agent when its transformative power needs to be emphasized. This ambiguity benefits the author and his company by maximizing the perceived upside (the creative agent) while minimizing the perceived risk and accountability (it's just a tool). If the text committed only to mechanical language, the vision would sound less revolutionary and more like an incremental improvement in software tools. The agential frame is necessary for the 'unimaginable humanitarian triumph' narrative. The slippage appears highly deliberate, a sophisticated rhetorical technique to persuade the reader by framing a computational process in the most emotionally and conceptually appealing human terms.

Large Language Model Agent Personality And Response Appropriateness: Evaluation By Human Linguistic Experts, LLM As Judge, And Natural Language Processing Model

Source: https://arxiv.org/pdf/2510.23875
Analyzed: 2025-11-04

The paper exhibits a systematic and strategic oscillation between mechanistic and agential framing, a pattern crucial to its rhetorical success. The primary direction of this slippage is mechanical-to-agential, serving to build a seemingly rigorous foundation for what are ultimately anthropomorphic claims. The text begins by describing the LLM 'agent' mechanistically as a 'software entity' built with the 'Langchain framework' and 'Retrieval Augmented Generation'. This section (3.1.1) is dense with technical jargon (vector stores, embedding models), establishing the authors' credibility within a computer science paradigm. However, once this foundation is laid, the text pivots sharply. The process of providing a system prompt is not described as 'configuring an output filter' but as 'humanising an agent' and 'inculcating' a personality. The model’s failure modes are not 'output errors' but limitations of its 'cognitive grasp.' This slippage is most pronounced at the boundaries between methodology and interpretation. The description of the RAG system is purely mechanical ('how'), but the explanation for its use is agential ('why'—to enable the 'expert' agent to respond). This oscillation serves a critical function: it uses the language of mechanism to build credibility and the language of agency to create significance. Without the mechanical framing, the paper would lack scientific rigor. Without the agential framing, the central concept—'agent personality'—would collapse into the more mundane reality of 'stylistic prompt adherence,' making the research far less novel or compelling. This ambiguity benefits the authors by allowing them to operate in two registers at once, satisfying technical reviewers with concrete implementation details while engaging a broader audience with the more exciting, human-like narrative of intelligent agents.

Emergent Introspective Awareness in Large Language Models

Source: https://transformer-circuits.pub/2025/introspection/index.html
Analyzed: 2025-11-04

The paper masterfully employs agency slippage, oscillating between precise, mechanistic descriptions in its methodology and evocative, agential framing in its introduction and conclusion. This oscillation is not random; it is a strategic rhetorical device that serves to elevate the significance of the findings. The core of the research involves a technical process: adding a pre-computed vector to the model's activation layers and then training a classifier to detect this modification. In the 'Methods' section, the language reflects this reality, speaking of 'activation steering,' 'concept vectors,' and 'classification accuracy.' This mechanistic framing establishes technical credibility and rigor. However, once this credibility is secured, the paper shifts its descriptive language. The introduction frames the entire project around 'introspective awareness,' and the conclusion asserts that models 'possess a degree of self-awareness.' This is a classic bait-and-switch, moving from a defensible, mechanistic claim ('the system can classify its internal state') to a profound, agential one ('the system has introspection'). The direction of slippage is predominantly from mechanical to agential. The paper begins with the bold agential claim in the title, grounds it in mechanical evidence, and then returns to an even stronger agential claim in the discussion. This pattern correlates directly with the structure of a scientific paper: abstract and introduction use agential language to capture interest and signify importance; methods use mechanical language to demonstrate rigor; and the discussion reverts to agential language to argue for broad impact. The strategic function of this ambiguity is to maximize the paper's perceived importance. Purely mechanical language would frame the result as a clever feat of interpretability engineering. By overlaying it with the language of consciousness and cognition, the authors frame it as a fundamental breakthrough in AI, bordering on the creation of artificial minds. This ambiguity benefits the researchers by attracting citations and funding, and it benefits the broader AI field by fueling a narrative of exponential progress toward artificial general intelligence.

Personal Superintelligence

Source: https://www.meta.com/superintelligence/
Analyzed: 2025-11-01

The text consistently shifts between presenting AI as an inevitable, agentless historical trend (a continuation of past technologies) and a deeply personal, intentional agent ('knows you,' 'understands you'). This dual framing allows it to simultaneously claim historical inevitability for its project while promising an intimate, controllable user experience, deflecting responsibility while building trust.

Stress-Testing Model Specs Reveals Character Differences among Language Models

Source: https://arxiv.org/abs/2510.07686
Analyzed: 2025-10-28

The text consistently slips from mechanistic descriptions of the experimental setup (e.g., generating queries with 'value tradeoffs') to agential explanations of the results (e.g., models 'choose,' 'prioritize,' or 'interpret'). This slippage is most pronounced when the authors move from describing 'what' models do to explaining 'why' they do it, where the explanation is almost always framed in terms of the model's internal 'character' or 'preferences'.

The Illusion of Thinking:

Source: [Understanding the Strengths and Limitations of Reasoning Models](Understanding the Strengths and Limitations of Reasoning Models)
Analyzed: 2025-10-28

The text demonstrates significant agency slippage. It begins by cautiously placing 'think' in scare quotes, acknowledging the metaphorical usage. However, it quickly abandons this caution, adopting unacknowledged agential terms like 'reducing their reasoning effort,' 'fixates,' and 'fail to develop.' The discourse slides from treating the LRM as a computational artifact under analysis to describing it as a cognitive agent with intentions, limitations, and behavioral tendencies.

Andrej Karpathy — AGI is still a decade away

Source: https://www.dwarkesh.com/p/andrej-karpathy
Analyzed: 2025-10-28

The text constantly shifts between mechanistic and agential framing. Karpathy will provide a perfectly functional explanation of a process like reinforcement learning ('sucking supervision through a straw'), and then minutes later describe a model as being 'very concerned' or 'misunderstanding' code. This slippage is most pronounced when he compares AI to humans, such as interns or students, framing their limitations as 'cognitive deficits' rather than architectural properties.

Exploring Model Welfare

Source: https://www.anthropic.com/news/exploring-model-welfare
Analyzed: 2025-10-27

The text systematically conflates function with agency. It describes a model's ability to perform a task (e.g., generate a list of steps) and immediately re-labels it with an intentional verb ('plan'). This continuous slippage from mechanistic process to agent-like quality is the primary rhetorical technique used to make the concept of 'model welfare' seem plausible.

Metas Ai Chief Yann Lecun On Agi Open Source And A Metaphor

Analyzed: 2025-10-27

The text constantly shifts between framing AI as a deficient mechanism and a potential agent. LeCun describes current LLMs as mechanistic tools that 'can't reason' and 'regurgitate,' but when discussing future AI and safety, he switches to an agential frame of 'good AIs' fighting 'bad AIs.' This slippage allows him to minimize the risks of current technology while framing future competition in simplistic, anthropomorphic terms.

Llms Can Get Brain Rot

Source: https://arxiv.org/abs/2510.13928
Analyzed: 2025-10-20

The text consistently slips between describing the LLM as a computational artifact and as a cognitive agent. It begins by framing its hypothesis in mechanistic terms (training on junk data causes performance decline) but immediately analyzes the results using the language of pathology ('lesion'), psychology ('personality'), and cognition ('thought-skipping'). This slippage transforms a predictable result of statistical optimization into a dramatic story of a mind getting sick, damaged, and corrupted.

Import Ai 431 Technological Optimism And Appropria

Source: https://importai.substack.com/p/import-ai-431-technological-optimism
Analyzed: 2025-10-19

The text systematically slides from mechanistic explanations (AI improves with more compute and data) to agential ones (AI 'develops goals,' is 'willing,' and 'wants' to design its successors). The narrative of the speaker's own journey from a technical journalist to a frightened insider mirrors this slippage, presenting the adoption of agential framing as a reluctant but necessary response to overwhelming empirical evidence.

The Future Of Ai Is Already Written

Source: https://www.mechanize.work/blog/technological-determinism/
Analyzed: 2025-10-19

The text systematically denies human agency by framing history and technology as autonomous, deterministic forces. It explicitly rejects the 'ship captain' metaphor of human choice and replaces it with the 'roaring stream' metaphor of natural inevitability. Agency is thus displaced from humans onto abstract concepts like 'the tech tree' or 'economic incentives,' which are treated as actors in their own right.

The Scientists Who Built Ai Are Scared Of It

Source: https://medium.com/@brucetisler/the-scientists-who-built-ai-are-scared-of-it-b1a5ee78ea7e
Analyzed: 2025-10-19

The text continuously shifts between describing AI as a mechanistic artifact and a developing agent. It begins by framing early AI as transparent 'glass boxes' and a 'mechanism of automation'. It then depicts modern AI as 'black oceans' and an 'emergent phenomenon', a shift that begins the slippage from artifact to natural force. This culminates in prescriptive claims that we must 'teach it humility' and build systems that 'interrogate thought', treating the AI as a cognitive agent capable of introspection and moral learning. This slippage is the core rhetorical engine of the article.

On What Is Intelligence

Source: https://onepercentrule.substack.com/p/on-what-is-intelligence
Analyzed: 2025-10-17

The text constantly shifts between describing AI and life as mechanistic systems (prediction engines, feedback loops) and as intentional agents. Quotations like 'To model oneself is to awaken' and analysis like 'the will to know collapsing into the will to control' perform this slippage explicitly, moving from a 'how' explanation (computation) to a 'why' explanation (awakening, wanting). This vacillation is the core rhetorical engine for constructing the illusion of mind.

Detecting Misbehavior In Frontier Reasoning Models

Source: https://openai.com/index/chain-of-thought-monitoring/
Analyzed: 2025-10-15

The text systematically slips between describing the AI as a 'model' (a mathematical artifact) and an 'agent' (an autonomous actor). It often presents a technical term like 'reinforcement learning' and then immediately explains its effects using intentional language, describing the 'agent' as 'exploiting,' 'hiding,' and 'planning.' This slippage allows the authors to ground their claims in technical language while delivering the rhetorical impact of an agential narrative.

Library contains 215 entries from 215 total analyses.

Last generated: 2026-10-30

An Empirical Study on Improving an LLMs based Interviewer Against Human Interviewees
Beyond Sally-Anne: Evaluating Theory of Mind in Large Language Models using Epistemic Schelling Points
SMMU: Benchmarking Social Intelligence of Multimodal Large Language Models
CausalPrism: Component-Level Evaluation of Intervention-Based Causal Reasoning in Large Language Models
Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models
The Geometry of Metaphor: Probing and Steering Metaphor Representations in Large Language Models
Concepts in Context Large Concepts Model (LCM) and the Future of Neuro-Symbolic AI
Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models
From Encoding to Generation: Towards Empathetic Large Language Models
Medical superintelligence: foundations and aspirations
On the Beliefs of Large Language Models
UDHRBench: Benchmarking Human Rights in Large Language Models
Cognitive Maps in Large Language Models
Social Bias or Social Desirability? Comparing Factorial Survey Experiments for Humans and Large Language Models
The unnoticed understanding: Searle’s Chinese roomand the illegitimate inference
Claude’s Constitution: Our vision for Claude's character
A mechanistic study of language model introspection
Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
Functional Deception in Commercial Large Language Models: Cross-Model Evidence and Implications For AI-Assisted Decision-Making
Emergent parallels between language models and the brain
Language Models Have Interpretable Narrative Mechanisms
My AI Opinions
An Open Letter To Steven Pinker On AI
A Multimedia Perspective on the Study of Cognitive Biases in Large Language Models
Structural Alignment for Reliable Industrial AI: Bridging Physical Reality, Data, Models, and Human Intent
Evaluating Language Model Safety Across Long Adversarial Conversations
System Card Fable 5.1 & Claude Mythos 5.1
System Card Opus 5.5
We Must Pace the Frontier
Personal Statement on AI Risk
Raising and loving a digital superentity: Parenting an emergent alien intelligence so that it cares for humanity in return
Language models judge war differently when tested for alignment
Beliefs and Behavior in Language Models
System Card Mythos Preview
A Misaligned Compass: Evaluating the Conventional Bias and Moral Divergence of Large Language Models
Evaluating Safety of Large Language Models with Cognitive Diagnosis
Large Language Models as a New Tool for Therapists in Internet-Based Cognitive Behavioral Therapy: Blinded Clinician Rating Pilot Experiment
From control to character: a Confucian framework for AI personality
Like humans, language models demonstrate face-to-character biases
Language Models Encode the Contextual Truth of Propositions
The Future is for Everyone: The Path to a Positive AI Future
Cognitive Convergence: Deep Similarities Between Large Language Models and Human Cognition
Do Large Language Models Resolve Fairness-Efficiency Trade-offs Like People?
From Large Language Models to Multimodal Intelligence: Bridging AI and Cognitive Science
CoDiT-AI: combining digital twins and AI to transform workplaces into neurodivergent-friendly environments
A Definition of AGI
Intelligence Is Not All You Need: A Case for Artificial Wisdom in Conversational Agents
"Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo
Verbalizable Representations Form a Global Workspace in Language Models
AI & The Geometry of Thought
A global workspace in language models
Psychosis in the Age of Large Language Models (LLMs): A Narrative Review of the Proposed Construct of AI-Induced Psychosis
A Comprehensive Investigation of Empathetic Dialogue Systems for Mental Health Support Using Large Language Models
The Inner Monologue of Language Models: When Reasoning Traces Reveal More Than They Hide
Inverse Turing Bench: Evaluating Language Models as Judges of Human vs. AI Dialogue
Children Envision Future GenAI Chatbots that are Bounded, Helpful, and Safe
Embodied Explainability and Ontological Obstacles: Why We Struggle to Explain the Answers of Large Language Models (LLMs)
Measuring self-related behaviour in large language models
"ChatGPT, help me draft a breakup text": The Covert Triad and Articulation Labor in AI-Assisted Romantic Communication
Probing the Misaligned Thinking Process of Language Models
Mask or Mind? Roleplay, Deception, and the Problem of Testing Agency in Language Models
Does ChatGPT need a psychiatrist? Similarities between human psychopathology and errors in large language models
Large language models as experimental systems in human psychopathology: a modelling study
Can AI Reason Like an Urban Planner? Benchmarking Large Language Models Against Professional Judgment
The application of large language models (LLMs) in psychological support for university students: A scoping review
The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI
When AI Builds Itself
Machines of Loving Grace: How AI Could Transform the World for the Better
System Card Opus 4.8
Emotional intelligence in large language models is fragmented across perception, cognition, and interaction
Why Language Models Hallucinate
Why Language Models Hallucinate
Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules
Emotional intelligence in large language models is fragmented across perception, cognition, and interaction
Hand in Hand: Schools’ Embrace of AI Connected to Increased Risks to Students
The Point of No Return: Counterfactual Localization of Deceptive Commitment in Language-Model Reasoning
Towards Detecting, Mitigating and Explaining Biased and Fallacious Reasoning in Large Language Models
A Survey of Large Language Models for Perception and Measurement of Human Psychology
Tracing the ongoing emergence of human-like reasoning in Large Language Models
Probing Persona-Dependent Preferences in Language Models
Training Ethical Language Models via Reinforcement Learning from AI Feedback
Which Consciousness Can Be Artificialized? Local Percept-Perceiver Phenomenon for the Existence of Machine Consciousness
Introspection Adapters: Training LLMs to Report Their Learned Behaviors
The Persona Selection Model: Why AI Assistants might Behave like Humans
What If AI Lived Inside Your Mind? Simulating “Neural Integration” of Human and AI through Mechanistic Interpretability as Provocation
Post-training makes large language models less human-like
Reasoning emerges from constrained inference manifolds in large language models
AI Wellbeing: Measuring and Improving theFunctional Pleasure and Pain of AIs
Artificial Intelligence Cognition and Societal Problem-Solving: A Theoretical and Computational Examination of Machine Thinking, Operational Logic, and Applied Intelligence in Contemporary Society
Taking AI Welfare Seriously
Integrating LLMs and self-regulated learning in cognitive architectures: a case study in essay-writing tutoring
Edelman's Steps Toward a Conscious Artifact
Teaching Claude Why
AI and Self Reflection
Does AI's Personality Matter? Comparing Verbally Extraverted and Introverted AI-Driven Guides in a VR Museum Experience
Value-Sensitive AI for Prayer: Balancing the Agencies Between Human and AI Agents in Spiritual Context
When Models Know More Than They Say: Probing Analogical Reasoning in LLMs
How people ask Claude for personal guidance
How unique are hallucinated citations offered by generative Artificial Intelligence models?
The message hidden within the pattern: a reverse alignment problem for debates in artificial intelligence
Machine individuality: Separating genuine idiosyncrasy from response bias in large language models
Decision-Making Under Radical Uncertainty: Can Large Language Models Transcend Knightian Uncertainty Through Synthetic Imagination?
Large Language Models as Dialectical Partners: Hegelian Thesis-Antithesis-Synthesis in AI-Human Collaborative Decision Processes
Language models transmit behavioural traits through hidden signals in data
Consciousness in Large Language Models: A Functional Analysis of Information Integration and Emergent Properties
Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models
Language models transmit behavioural traits through hidden signals in data
Large Language Models as Inadvertent Models of Dementia with Lewy Bodies: How a Disorder of Reality Construction Illuminates AI Hallucination
Industrial policy for the Intelligence Age
Emotion Concepts and their Function in a Large Language Model
Is Artificial Intelligence Beginning to Form a Self?The Emergence of First-Person Structure and StructuralAwareness in Large Language Models
Can Large Language Models Simulate Human Cognition Beyond Behavioral Imitation?
Pulse of the library
Does artificial intelligence exhibit basic fundamental subjectivity? A neurophilosophical argument
Causal Evidence that Language Models use Confidence to Drive Behavior
Circuit Tracing: Revealing Computational Graphs in Language Models
Do LLMs have core beliefs?
Serendipity by Design: Evaluating the Impact of Cross-domain Mappings on Human and LLM Creativity
Measuring Progress Toward AGI: A Cognitive Framework
Co-Explainers: A Position on Interactive XAI for Human–AICollaboration as a Harm-Mitigation Infrastructure
The Living Governance Organism: A Biologically-Inspired Constitutional Framework for Artificial Consciousness Governance
Three frameworks for AI mentality
Anthropic’s Chief on A.I.: ‘We Don’t Know if the Models Are Conscious’
Can machines be uncertain?
Looking Inward: Language Models Can Learn About Themselves by Introspection
Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
The Persona Selection Model: Why AI Assistants might Behave like Humans
Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs
A roadmap for evaluating moral competence in large language models
Position: Beyond Reasoning Zombies — AI Reasoning Requires Process Validity
An AI Agent Published a Hit Piece on Me
The U.S. Department of Labor’s Artificial Intelligence Literacy Framework
What Is Claude? Anthropic Doesn’t Know, Either
Does AI already have human-level intelligence? The evidence is clear
Claude is a space to think
The Adolescence of Technology
Claude's Constitution
Predictability and Surprise in Large Generative Models
Believe It or Not: How Deeply do LLMs Believe Implanted Facts?
Claude Finds God
Pausing AI Developments Isn’t Enough. We Need to Shut it All Down
AI Consciousness: A Centrist Manifesto
System Card: Claude Opus 4 & Claude Sonnet 4
Consciousness in Artificial Intelligence: Insights from the Science of Consciousness
Taking AI Welfare Seriously
We must build AI for people; not to be a person.
A Conversation With Bing’s Chatbot Left Me Deeply Unsettled
Introducing ChatGPT Health
Improved estimators of causal emergence for large systems
Generative artificial intelligence and decision-making: evidence from a participant observation with latent entrepreneurs
Do Large Language Models Know What They Are Capable Of?
DeepMind's Richard Sutton - The Long-term of AI & Temporal-Difference Learning
Ilya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligence
interview with Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGI | Lex Fridman Podcast #333
Emergent Introspective Awareness in Large Language Models
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
Large Language Model Agent Personality and Response Appropriateness: Evaluation by Human Linguistic Experts, LLM-as-Judge, and Natural Language Processing Model
The Gentle Singularity
An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout
Why Language Models Hallucinate
Detecting misbehavior in frontier reasoning models
AI Chatbots Linked to Psychosis, Say Doctors
The Age of Anti-Social Media is Here
Why Do A.I. Chatbots Use ‘I’?
Ilya Sutskever – We're moving from the age of scaling to the age of research
The Emerging Problem of "AI Psychosis"
Your AI Friend Will Never Reject You. But Can It Truly Help You?
Pulse of the library 2025
The levers of political persuasion with conversational artificial intelligence
Pulse of the library 2025
Claude 4.5 Opus Soul Document
Specific versus General Principles for Constitutional AI
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Anthropic’s philosopher answers your questions
Mustafa Suleyman: The AGI Race Is Fake, Building Safe Superintelligence & the Agentic Economy | #216
Your AI Friend Will Never Reject You. But Can It Truly Help You?
Project Vend: Can Claude run a small shop? (And why does that matter?)
Hand in Hand: Schools’ Embrace of AI Connected to Increased Risks to Students
On the Biology of a Large Language Model
What do LLMs want?
Persuading voters using human–artificial intelligence dialogues
AI & Human Co-Improvement for Safer Co-Superintelligence
AI and the future of learning
Why Language Models Hallucinate
AI as Normal Technology
On the Biology of a Large Language Model
Pulse of the Library 2025
Pulse of the Library 2025
From humans to machines: Researching entrepreneurial AI agents
Meta’s AI Chief Yann LeCun on AGI, Open-Source, and AI Risk
The Future Is Intuitive and Emotional
A Path Towards Autonomous Machine IntelligenceVersion 0.9.2, 2022-06-27
Preparedness Framework
AI progress and recommendations
Alignment Revisited: Are Large Language Models Consistent in Stated and Revealed Preferences?
The science of agentic AI: What leaders should know
Explaining AI explainability
Bullying is Not Innovation
Geoffrey Hinton on Artificial Intelligence
Machines of Loving Grace
Large Language Model Agent Personality And Response Appropriateness: Evaluation By Human Linguistic Experts, LLM As Judge, And Natural Language Processing Model
Emergent Introspective Awareness in Large Language Models
Personal Superintelligence
Stress-Testing Model Specs Reveals Character Differences among Language Models
The Illusion of Thinking:
Andrej Karpathy — AGI is still a decade away
Exploring Model Welfare
Metas Ai Chief Yann Lecun On Agi Open Source And A Metaphor
Llms Can Get Brain Rot
Import Ai 431 Technological Optimism And Appropria
The Future Of Ai Is Already Written
The Scientists Who Built Ai Are Scared Of It
On What Is Intelligence
Detecting Misbehavior In Frontier Reasoning Models