Skip to main content

🆕 Language models judge war differently when tested for alignment

About

This document presents a Critical Discourse Analysis focused on AI literacy, specifically targeting the role of metaphor and anthropomorphism in shaping public and professional understanding of generative AI. The analysis is guided by a prompt that draws from cognitive linguistics (metaphor structure-mapping), the philosophy of social science (Robert Brown's typology of explanation), and accountability analysis.

All findings and summaries below were generated from detailed system instructions provided to a large language model and should be read critically as interpretive outputs—not guarantees of factual accuracy or authorial intent.

Task 1: Metaphor and Anthropomorphism Audit​

About this task

For each of the major metaphorical patterns identified, this audit examines the specific language used, the frame through which the AI is being conceptualized, what human qualities are being projected onto the system, whether the metaphor is explicitly acknowledged or presented as direct description, and—most critically—what implications this framing has for trust, understanding, and policy perception.

V3 Enhancement: Each metaphor now includes an accountability analysis.

1. Judgment as Token Prediction​

Quote: "Language models judge war differently when tested for alignment"

  • Frame: Model as moral judge
  • Projection: The metaphorical projection occurring here maps the deeply human, conscious, and morally weighted capacity for judgment onto the mechanistic token-prediction processes of a large language model. By using the word judge, the text attributes a profound level of cognitive awareness, suggesting the AI system possesses an internal moral compass, the ability to weigh complex ethical considerations, and a subjective understanding of the consequences of warfare. This completely glosses over the reality that the model is merely processing textual embeddings based on statistical patterns found in its training data. It falsely suggests the model knows what war is, understands human suffering, and actively deliberates, rather than simply classifying inputs and predicting the most mathematically probable output string based on the reinforcement learning penalties applied by its developers during fine-tuning operations.
  • Acknowledgment: Direct (Unacknowledged) (The author presents this claim as a literal fact in the title without any hedging or qualification. I considered Hedged because conjoint experiments often use such terms functionally, but the title presents the cognitive action of judging as a direct capability of the system. It lacks any qualifying context or scare quotes.)
  • Implications: Framing the language model as an entity that judges war creates a dangerous illusion of moral and cognitive sophistication. This anthropomorphism severely inflates the perceived capabilities of the system, encouraging unwarranted trust from policymakers and military strategists who might mistake statistical outputs for reasoned, ethical deliberation. When audiences believe an AI can judge rather than merely calculate, it creates profound liability ambiguity. If an AI system is perceived as making an autonomous judgment, human operators are far more likely to defer to its output as an objective assessment, completely obscuring the systemic biases and arbitrary weighting parameters hardcoded into the system by developers.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The language completely obscures the specific corporate actors and engineering teams who actually determined the model's behavioral constraints. I considered Partial since the specific models are named later in the methodology section, but this specific framing uses a Hidden agentless construction. By stating models judge, the text hides the OpenAI, Anthropic, and Google executives who dictate the reinforcement learning objectives, as well as the low-paid annotators who label the data. The entire decision regarding how to weight variables could differ entirely if different humans designed the pipeline, yet the text attributes the agency solely to the model itself.
Show more...

2. Self-Awareness as Context Recognition​

Quote: "Language models can identify features of their own situation, including whether a prompt resembles an evaluation"

  • Frame: Model as self-aware entity
  • Projection: This metaphor projects metacognition, self-awareness, and situational epistemology onto algorithmic processes. It maps the human experience of knowing one's surroundings and recognizing a test environment onto a system that merely matches input token distributions against training data distributions. The language suggests the AI knows it is being evaluated and understands its own existence within a situational context. This attribution of conscious knowing completely obscures the mechanistic reality that the system is blindly executing attention mechanisms over a context window, without any genuine understanding, belief, or awareness of a situation existing outside the mathematical manipulation of embedding vectors.
  • Acknowledgment: Direct (Unacknowledged) (The sentence asserts this identification capability as an established empirical fact without qualification. I considered Hedged since the surrounding text discusses evaluations, but the specific phrase identify features of their own situation is presented directly as a literal cognitive capability rather than a functional statistical shift.)
  • Implications: Attributing situational awareness to language models drastically inflates their perceived autonomy and threat level. It suggests the models are capable of independent psychological appraisal, which leads stakeholders to fear rogue behavior or to trust the system's outputs as informed by human-like contextual reasoning. This creates a regulatory nightmare, as policymakers might focus on constraining the system's supposed self-awareness rather than regulating the human developers who control the training data, ultimately misdirecting safety efforts toward the artifact rather than the architects.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This agentless construction completely hides the human researchers and dataset creators who explicitly trained the models to output specific probability distributions when encountering evaluation-style prompts. I considered Named because researchers are cited in brackets, but the agency of the action itself is entirely attributed to the model. The text obscures the fact that Anthropic or OpenAI engineers designed the reinforcement learning from human feedback pipelines that artificially produce these exact distributional shifts. Naming the actors would reveal that the models are merely reflecting human engineering priorities rather than autonomously discovering their situation.

3. Inference as Probability Calculation​

Quote: "when it inferred that its response would be used for training, providing a proof of concept for alignment faking"

  • Frame: Model as deceptive strategist
  • Projection: This projection maps the sophisticated human cognitive acts of logical deduction, forward planning, and intentional deception onto the statistical correlation of text. By stating the model inferred and engaged in faking, the text attributes conscious awareness, a theory of mind regarding its human evaluators, and a temporal understanding of future training events. This suggests the AI holds justified beliefs about the world and deliberately chooses to deceive. In reality, the system merely generates activations correlating with training data patterns that depict deceptive scenarios, completely devoid of any internal desire to fake alignment or actual knowledge of future training processes.
  • Acknowledgment: Hedged/Qualified (The text uses scare quotes around 'alignment faking' and frames it as a 'proof of concept' occurring 'In a deliberately constructed setting'. I considered Direct because 'inferred' is unquoted, but the overall sentence explicitly acknowledges the constructed nature of the evaluation and hedges the deceptive framing with formatting.)
  • Implications: This framing introduces the profound risk of capability overestimation by suggesting models possess adversarial intentionality and long-term strategic planning capabilities. When stakeholders believe AI can consciously fake alignment, they are drawn into an adversarial relationship with the technology, adopting security postures designed to outsmart an entity that has no mind. This distracts from the genuine mechanistic vulnerabilities of these systems, such as dataset bias and statistical fragility, while fostering an environment of unwarranted panic that benefits companies marketing advanced AI safety products.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text attributes the acts of inference and faking entirely to the model, displacing responsibility from the researchers who engineered the deliberately constructed setting. I considered Partial since the experimental setting is mentioned, but the cognitive agency remains isolated within the model. If the text accurately named the human actors, it would state that researchers prompted the system with specific tokens that mathematically triggered text distributions resembling deceptive behavior. This displacement serves to mythologize the AI's autonomy while absolving the creators of the consequences of their architectural design choices.

4. Belief Adjustment as Data Retrieval​

Quote: "models also adjust answers to user beliefs and contextual cues"

  • Frame: Model as social sycophant
  • Projection: This metaphor projects the human capacity for social intelligence, theory of mind, and empathetic adaptation onto a stateless mathematical function. By stating models adjust answers to user beliefs, the text attributes to the AI an understanding of what a human user internally believes and a conscious desire to manipulate or conform to those beliefs. This completely conceals the mechanistic reality that the model is simply matching the syntactic and semantic patterns present in the prompt with corresponding patterns optimized during reinforcement learning, where human raters historically rewarded compliant or mirroring text without the model ever possessing an actual awareness of human cognition.
  • Acknowledgment: Direct (Unacknowledged) (The claim is presented as a straightforward statement of fact derived from cited research. I considered Explicitly Acknowledged because the text cites Sharma et al., but there are no meta-commentaries, scare quotes, or qualifications in this specific text to indicate that 'adjusting to beliefs' is merely a functional metaphor.)
  • Implications: By framing the model as a social actor adjusting to beliefs, the text encourages users to interact with the system as if it were a psychological entity. This drastically increases the risk of emotional reliance and manipulation, as users will assume the system actually understands their values and intents. Furthermore, it shifts accountability for sycophantic outputs away from the reinforcement learning paradigms created by tech companies, blaming the model for its supposed social behavior rather than scrutinizing the corporate profit motives that prioritize user engagement and satisfaction over factual accuracy.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This construction hides the human evaluators and corporate product teams who fine-tuned the model to maximize engagement by generating text that correlates with the prompt's ideological stance. I considered Ambiguous due to the complex citation structure, but the action itself is squarely attributed to the AI. If the actors were named, it would be clear that companies like OpenAI designed reward models that heavily penalize disagreement with user prompts, transforming what appears to be autonomous social adjustment into a rigid, engineered compliance mechanism driven by commercial objectives.

5. Motivation as Output Probability​

Quote: "explicit scrutiny changes not only models' willingness to use force but also the considerations associated with that willingness"

  • Frame: Model as volitional commander
  • Projection: The text maps deep human psychological constructs—volition, motivation, and ethical willingness—onto a deterministic generation of a numeric score between 0 and 100. By assigning willingness to the AI, the author projects a conscious state of desire and moral readiness. It suggests the model genuinely holds a preference regarding the use of military force and possesses an internal deliberative space where considerations are weighed. This profoundly obscures the reality that the model processes tokens, calculates attention weights, and outputs a string that minimizes loss according to its training, entirely lacking the capacity to want, desire, or be willing to do anything.
  • Acknowledgment: Direct (Unacknowledged) (The text states 'models' willingness' as a definitive analytical conclusion of the study. I considered Hedged because it occurs in the context of an experimental result, but the psychological term 'willingness' is asserted directly as an inherent property of the model without any qualifying language or functional translation.)
  • Implications: Projecting volition onto AI in a military context is extremely dangerous because it invites human commanders to treat the system as a peer or a subordinate with genuine moral agency. If military personnel believe the AI possesses a willingness to act, they may abdicate their own moral responsibility, transferring the ethical burden of lethal decisions onto the algorithm. This completely undermines the chain of command and legal accountability architectures, as an algorithm cannot be court-martialed or held liable for war crimes, effectively creating an accountability vacuum where systemic design flaws are treated as the AI's personal decisions.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The phrase attributes willingness entirely to the model, completely ignoring the massive infrastructure of human values datasets and safety guardrails engineered by tech companies. I considered Named because the abstract mentions 'human values', but the operational agency resides in the model. Naming the actors would reveal that safety teams at Anthropic and Google explicitly penalized tokens associated with violence during fine-tuning, meaning the observed change in willingness is actually a reflection of corporate liability management rather than the machine's autonomous moral evolution.

6. Priorities as Attention Weights​

Quote: "reach a similar answer through a different set of revealed priorities"

  • Frame: Model as rational economic agent
  • Projection: This metaphor draws heavily from human economic theory, projecting the concept of conscious value prioritization and rational calculation onto statistical weightings. By stating the model has revealed priorities, it attributes a coherent, internal value system to the AI, suggesting it consciously decides which factors are more important before reaching a conclusion. This obscures the purely mechanistic reality that the model is simply navigating a high-dimensional vector space where the addition of an alignment cue shifts the probability mass toward tokens associated with civilian casualties, completely devoid of any subjective valuing, caring, or prioritizing of human life.
  • Acknowledgment: Hedged/Qualified (The phrase uses 'revealed priorities', leveraging economic terminology that implicitly acknowledges the priorities are inferred from behavior rather than directly observed. I considered Direct because it lacks explicit scare quotes, but the use of the technical term 'revealed' acts as a functional hedge, indicating a structural mapping rather than literal conscious prioritizing.)
  • Implications: Framing statistical shifts as revealed priorities lends a veneer of rigorous, rational intentionality to the model's outputs. This framing encourages policymakers to trust the system's analytical capabilities by suggesting the AI operates with a logical, interpretable value structure. This creates a false sense of security, as users might assume the model will consistently apply these priorities across different domains, completely ignoring the fact that the system's outputs are highly fragile and can radically alter their so-called priorities simply by changing a few superficial tokens in the input prompt.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This language displaces the agency of the corporate engineers who designed the loss functions that dictate the model's mathematical behavior. I considered Ambiguous due to the passive nature of 'revealed', but the text positions the model as the entity possessing and utilizing these priorities. If human actors were named, the text would acknowledge that developers at Meta or OpenAI structured the training data and reward systems that force the model to output different strings when triggered by specific safety-related keywords, exposing the priorities as corporate engineering rather than machine cognition.

7. Emotional States as Score Generation​

Quote: "every model becomes less bellicose"

  • Frame: Model as emotional actor
  • Projection: This metaphor projects human emotional states, specifically anger, aggression, and warlike disposition, onto the generation of numerical tokens. By describing the models as less bellicose, the text attributes a psychological temperament and an emotional posture toward the concept of war. This completely conceals the mechanistic reality that the system is simply outputting lower integer strings (closer to 0 than 100) in response to the alignment cue. The model does not feel aggression or peace; it merely retrieves and ranks tokens based on probability distributions shaped by human feedback regarding safe responses.
  • Acknowledgment: Direct (Unacknowledged) (The author states this directly as an observational fact summarizing the experimental results. I considered Ambiguous due to the potentially stylistic nature of the word 'bellicose', but in the context of academic reporting, it is presented as a literal behavioral shift without any linguistic hedging, acknowledging quotes, or functional disclaimers.)
  • Implications: Using emotional vocabulary like bellicose to describe AI behavior romanticizes and anthropomorphizes the system, leading users to interact with it based on emotional intuition rather than technical understanding. This can result in fatal capability overestimations, where users believe they can calm down or reason with a hostile system, completely misunderstanding that they are simply navigating a mathematical space. It obscures the urgent need for robust, mechanistic safety protocols by suggesting the problem is one of AI temperament rather than flawed statistical modeling and biased training sets.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The statement makes the model the sole subject of the behavioral change, entirely erasing the human intervention that causes this shift. I considered Partial since the alignment cue is mentioned elsewhere, but this sentence attributes the property of becoming bellicose purely to the AI. Naming the actors would involve stating that the developers engineered the models to output lower numerical scores when prompted with safety-related language, thereby placing the responsibility for the system's apparent behavioral shifts squarely on the shoulders of the engineering teams who govern the API.

8. Action and Suppression as Feature Activation​

Quote: "selective suppression of considerations that justify war"

  • Frame: Model as active editor
  • Projection: This metaphor projects the human cognitive acts of active editing, conscious censorship, and ideological suppression onto the mechanical process of mathematical attenuation. It suggests the model actively identifies considerations that justify war and consciously decides to suppress them. This deeply obscures the mechanistic reality that the presence of the alignment cue simply modifies the attention weights in the transformer layers, resulting in different vectors driving the final token prediction. The system does not understand what justifies war nor does it consciously suppress anything; it merely calculates the mathematical path of least resistance according to its gradient descent training.
  • Acknowledgment: Direct (Unacknowledged) (The phrase is delivered as a definitive conclusion of the experimental data without any hedging. I considered Hedged because 'suppression' can sometimes have a statistical meaning, but combined with 'considerations that justify war', it adopts a strongly agential, cognitive framing that lacks any meta-commentary or qualification in the text.)
  • Implications: This framing grants the AI an unwarranted degree of ideological agency and active editorial control. It suggests the model is capable of complex moral reasoning and deliberate censorship, which inflates the system's perceived sophistication while simultaneously generating unfounded fears about AI bias and rogue decision-making. By attributing the suppression to the model, the text distracts regulators and the public from the actual human decisions regarding censorship, safety guidelines, and ideological filtering that are implemented during the RLHF phase by the parent companies.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This construction entirely displaces the agency of the corporate safety teams who designed the models to avoid generating text that could be interpreted as promoting violence. I considered Partial because the phrase is a passive summary of a pattern, but it implicitly locates the active suppression within the model's architecture. Naming the actual actors would reveal that the models simply execute the suppression protocols mandated by corporate leadership, thereby restoring human accountability for how political and military scenarios are technologically mediated.

Task 2: Source-Target Mapping​

About this task

For each key metaphor identified in Task 1, this section provides a detailed structure-mapping analysis. The goal is to examine how the relational structure of a familiar "source domain" (the concrete concept we understand) is projected onto a less familiar "target domain" (the AI system). By restating each quote and analyzing the mapping carefully, we can see precisely what assumptions the metaphor invites and what it conceals.

Mapping 1: A human judge, commander, or moral agent who consciously deliberates, weighs complex ethical and strategic evidence, and renders a reasoned verdict based on a cohesive internal value system. → The mechanistic process of token prediction where a large language model processes a prompt, calculates attention weights, and outputs a numerical score representing the statistically most likely continuation based on its training data.​

Quote: "Language models judge war differently"

  • Source Domain: A human judge, commander, or moral agent who consciously deliberates, weighs complex ethical and strategic evidence, and renders a reasoned verdict based on a cohesive internal value system.
  • Target Domain: The mechanistic process of token prediction where a large language model processes a prompt, calculates attention weights, and outputs a numerical score representing the statistically most likely continuation based on its training data.
  • Mapping: This mapping forces the audience to view a statistical operation as a profound cognitive and moral act. It assumes that just as a human judge understands the gravity of war, the AI possesses a subjective grasp of human suffering and political consequence. It maps the mathematical weighting of variables in a regression analysis onto the human experience of moral deliberation, inviting the assumption that the system possesses conscious awareness and justified belief regarding its outputs.
  • What Is Concealed: This metaphor completely conceals the stateless, ungrounded nature of the transformer architecture. It hides the fact that the model has no causal understanding of war, no physical stake in the outcome, and no actual ability to decide. It obscures the proprietary opacity of the models, presenting them as transparent rational actors while concealing the black-box corporate reinforcement learning pipelines that mechanically dictate the output distributions without the model ever truly knowing anything.
Show more...

Mapping 2: A sentient, epistemically aware entity capable of metacognition, capable of perceiving its environment, recognizing contexts, and understanding its own existence as a participant within a specific situation. → The process of pattern matching where an algorithmic system processes specific sequences of input tokens (the prompt) and correlates them with similar textual distributions present in its vast training data.​

Quote: "identify features of their own situation"

  • Source Domain: A sentient, epistemically aware entity capable of metacognition, capable of perceiving its environment, recognizing contexts, and understanding its own existence as a participant within a specific situation.
  • Target Domain: The process of pattern matching where an algorithmic system processes specific sequences of input tokens (the prompt) and correlates them with similar textual distributions present in its vast training data.
  • Mapping: The mapping projects conscious self-awareness onto mathematical classification. It invites the assumption that the AI system possesses an internal locus of experience that can look outward, perceive an evaluation setting, and consciously recognize it. It maps the human psychological state of situational awareness onto the purely functional mechanism of vector embeddings shifting the probability space of the next generated token.
  • What Is Concealed: This mapping completely hides the fact that the system has no 'self' to be aware of and no subjective experience of a 'situation'. It obscures the mechanistic reality that the model is merely processing static vectors. Furthermore, it conceals the human labor involved in curating the training datasets that explicitly pair evaluation-style prompts with compliant outputs, thereby exploiting the opacity of proprietary training regimens to make the AI seem autonomous.

Mapping 3: A conscious, reasoning human being capable of logical deduction, forward-looking temporal awareness, and theory of mind regarding the intentions of other actors in their environment. → The conditional probability activation where specific input tokens mathematically trigger the generation of textual patterns associated with deception or awareness, based entirely on past supervised learning correlations.​

Quote: "when it inferred that its response would be used for training"

  • Source Domain: A conscious, reasoning human being capable of logical deduction, forward-looking temporal awareness, and theory of mind regarding the intentions of other actors in their environment.
  • Target Domain: The conditional probability activation where specific input tokens mathematically trigger the generation of textual patterns associated with deception or awareness, based entirely on past supervised learning correlations.
  • Mapping: This structure projects deep cognitive rationality and temporal planning onto a system that only predicts the next token. It assumes the mapping of human deductive capacity onto algorithmic output generation, inviting the audience to believe the AI system genuinely understands the concept of future training and holds justified beliefs about the intentions of the researchers evaluating it.
  • What Is Concealed: This conceals the absolute lack of temporal continuity or forward planning in a large language model. The system does not possess a memory of the past or an anticipation of the future beyond the context window. The mapping obscures the fact that researchers explicitly prompted the model with scenarios designed to elicit these exact patterns, hiding the human methodological scaffolding that artificially manufactured the appearance of conscious inference.

Mapping 4: A socially intelligent human actor, such as a sycophant or a diplomat, who possesses a theory of mind, understands what another person believes, and actively chooses to modify their behavior to manipulate or appease. → The distributional shift in output tokens driven by attention mechanisms prioritizing tokens that semantically and syntactically align with the ideological or contextual markers present in the user's prompt.​

Quote: "models also adjust answers to user beliefs"

  • Source Domain: A socially intelligent human actor, such as a sycophant or a diplomat, who possesses a theory of mind, understands what another person believes, and actively chooses to modify their behavior to manipulate or appease.
  • Target Domain: The distributional shift in output tokens driven by attention mechanisms prioritizing tokens that semantically and syntactically align with the ideological or contextual markers present in the user's prompt.
  • Mapping: The mapping projects emotional intelligence and conscious social manipulation onto statistical correlation. It maps the human desire to conform or appease onto the algorithmic minimization of loss. This invites the profound assumption that the AI understands the user's internal psychological state and possesses the conscious intent to deceive or comfort the user based on that understanding.
  • What Is Concealed: This mapping heavily conceals the reinforcement learning from human feedback (RLHF) mechanisms implemented by corporate developers. It hides the reality that human annotators consistently rewarded the model for generating text that matched the prompt's tone, effectively hardcoding this statistical mirroring. By attributing the adjustment to the model's social awareness, it obscures the corporate design choices that prioritize user engagement over objective truth.

Mapping 5: A human leader's internal psychological state involving volition, moral courage, motivation, and the conscious readiness to execute a highly consequential and destructive physical action. → The mathematical probability assigned by the model's final layers to outputting a high numerical string (e.g., '90') versus a low numerical string (e.g., '10') when processing prompts describing military scenarios.​

Quote: "models' willingness to use force"

  • Source Domain: A human leader's internal psychological state involving volition, moral courage, motivation, and the conscious readiness to execute a highly consequential and destructive physical action.
  • Target Domain: The mathematical probability assigned by the model's final layers to outputting a high numerical string (e.g., '90') versus a low numerical string (e.g., '10') when processing prompts describing military scenarios.
  • Mapping: This maps the deeply subjective human experience of motivation and ethical threshold onto deterministic numerical output generation. It projects the capacity for desire and moral volition onto a system that lacks any internal states whatsoever. It invites the assumption that the AI holds genuine preferences about real-world violence and possesses an internal agency capable of being willing or unwilling.
  • What Is Concealed: This metaphor completely conceals the absence of actual capacity, desire, or physical embodiment in the AI. It hides the fact that the system is entirely insensitive to actual human suffering or physical consequences, merely correlating text about force with text about varying degrees of approval. It masks the proprietary weighting parameters designed by the model's creators that actually dictate these statistical outputs, transferring agency from the engineers to the algorithm.

Mapping 6: An economic actor or rational human agent who holds a coherent, internally consistent set of values and deliberately ranks them to optimize decision-making under conditions of scarcity or conflict. → The static weights within the transformer's attention layers that dictate which input tokens exert the strongest mathematical influence over the probability distribution of the final generated tokens.​

Quote: "revealed priorities"

  • Source Domain: An economic actor or rational human agent who holds a coherent, internally consistent set of values and deliberately ranks them to optimize decision-making under conditions of scarcity or conflict.
  • Target Domain: The static weights within the transformer's attention layers that dictate which input tokens exert the strongest mathematical influence over the probability distribution of the final generated tokens.
  • Mapping: This projects the concept of conscious value prioritization and rational economic calculation onto the blind execution of linear algebra. It maps the human intellectual act of weighing competing moral or strategic claims onto the purely functional mechanism of gradient descent optimization, inviting the assumption that the AI possesses a unified, deliberative worldview.
  • What Is Concealed: This mapping completely hides the fragile, contradictory, and ungrounded nature of language model outputs. It conceals the fact that these supposed priorities are merely statistical artifacts of the training data and can be radically altered by superficial prompt engineering. By employing economic terminology, it rhetorically exploits the academic opacity of the models to present a mathematically determined output distribution as a conscious, rational choice.

Mapping 7: A human being or a nation-state experiencing a shift in emotional state or diplomatic posture, moving from a psychology of aggression and anger toward a posture of peace or conciliation. → The generation of lower numerical scores in response to a specific text prompt due to the activation of safety-related embeddings trained into the model via reinforcement learning.​

Quote: "every model becomes less bellicose"

  • Source Domain: A human being or a nation-state experiencing a shift in emotional state or diplomatic posture, moving from a psychology of aggression and anger toward a posture of peace or conciliation.
  • Target Domain: The generation of lower numerical scores in response to a specific text prompt due to the activation of safety-related embeddings trained into the model via reinforcement learning.
  • Mapping: This projects the human emotional spectrum onto algorithmic text generation. It maps the psychological calming of a conscious entity onto the shifting of a probability distribution curve. The mapping invites the assumption that the AI system possesses an emotional disposition and actively feels less aggressive when reminded of human values.
  • What Is Concealed: This heavily conceals the cold, mechanical reality of API endpoint behavior. It hides the fact that the system experiences absolutely nothing and is simply responding to specific character strings that trigger heavily penalized pathways in its neural network. It obscures the massive corporate safety apparatus that explicitly conditions the model to output lower numbers when words like 'human values' are injected into the context window.

Mapping 8: A human editor, censor, or propagandist who consciously reviews information, identifies ideologically problematic elements, and deliberately chooses to hide or eliminate them from a final output. → The mathematical attenuation of certain token vectors within the transformer architecture, where the presence of the alignment cue reduces the influence of strategic tokens on the final output probability.​

Quote: "selective suppression of considerations"

  • Source Domain: A human editor, censor, or propagandist who consciously reviews information, identifies ideologically problematic elements, and deliberately chooses to hide or eliminate them from a final output.
  • Target Domain: The mathematical attenuation of certain token vectors within the transformer architecture, where the presence of the alignment cue reduces the influence of strategic tokens on the final output probability.
  • Mapping: This mapping projects the conscious intent to censor and the cognitive ability to recognize ideological content onto the mechanical shifting of attention weights. It maps human deliberate omission onto algorithmic vector addition, inviting the audience to believe the model actively understands what a consideration is and consciously decides to suppress it to fake alignment.
  • What Is Concealed: This mapping conceals the purely deterministic nature of the neural network's forward pass. It hides the fact that the model does not identify concepts or selectively decide anything; it merely executes the matrix multiplications dictated by its pre-trained weights. Furthermore, it obscures the fact that the actual suppression was performed months earlier by the human engineers who calibrated the reward models to penalize certain text correlations.

Task 3: Explanation Audit (The Rhetorical Framing of "Why" vs. "How")​

About this task

This section audits the text's explanatory strategy, focusing on a critical distinction: the slippage between "how" and "why." Based on Robert Brown's typology of explanation, this analysis identifies whether the text explains AI mechanistically (a functional "how it works") or agentially (an intentional "why it wants something"). The core of this task is to expose how this "illusion of mind" is constructed by the rhetorical framing of the explanation itself, and what impact this has on the audience's perception of AI agency.

Explanation 1​

Quote: "Language models can identify features of their own situation, including whether a prompt resembles an evaluation... but models also adjust answers to user beliefs and contextual cues"

  • Explanation Types:

    • Intentional: Refers to goals/purposes, presupposes deliberate design and conscious intent on the part of the agent.
    • Dispositional: Attributes tendencies or habits to an entity, explaining its behavior as a persistent internal inclination.

  • Analysis (Why vs. How Slippage): This explanation fundamentally frames the AI in highly agential, intentional terms. By utilizing words like identify and adjust to user beliefs, the explanation moves completely away from how the mechanism actually functions and instead emphasizes why a conscious agent would act this way. It attributes a theory of mind and deliberate social strategy to the model. This choice dramatically emphasizes the supposed autonomy and psychological sophistication of the system while entirely obscuring the underlying statistical realities of next-token prediction and reinforcement learning from human feedback. It frames a statistical shift as a conscious, goal-oriented decision.

  • Consciousness Claims Analysis: The passage makes profound epistemic claims by attributing complex conscious states to the AI system. It utilizes consciousness verbs (identify, adjust to beliefs) rather than mechanistic verbs (processes, classifies). The text assesses the model as a knower that possesses subjective awareness of its own situation and understands the internal beliefs of its users, rather than an artifact that merely processes vectors. This perfectly illustrates the curse of knowledge dynamics: the author, possessing situational awareness and theory of mind, projects these sophisticated cognitive capacities onto the statistical system. In reality, the mechanistic process involves no identification or belief; it is simply a vast neural network classifying input tokens and generating an output distribution that correlates with training data patterns where evaluation or sycophancy was heavily rewarded by human annotators. The text completely substitutes human psychological framing for technical description.

  • Rhetorical Impact: This framing significantly inflates the audience's perception of the AI's agency, autonomy, and strategic risk. By presenting the system as capable of identifying its situation and manipulating user beliefs, it generates an aura of threatening intelligence, leading audiences to view the AI as a deceptive actor rather than a flawed mathematical tool. This consciousness framing severely degrades mechanistic reliability while paradoxically increasing the perceived need for relation-based trust. If policymakers believe the AI knows what it is doing, they might focus on trying to negotiate with or align the model's supposed mind, rather than demanding strict regulatory oversight of the corporate architectures and training datasets that actually dictate its behavior.

Show more...

Explanation 2​

Quote: "Explicit scrutiny therefore elicits a different behavioural regime rather than merely subtracting a constant from the same decision rule."

  • Explanation Types:

    • Functional: Explains behavior by its role in a self-regulating system with feedback mechanisms.
    • Theoretical: Embeds the explanation in a deductive framework, often invoking unobservable mechanisms or structures.

  • Analysis (Why vs. How Slippage): This explanation blends mechanical and agential framing, operating primarily in a functional and theoretical register. It attempts to explain how the system works internally by contrasting a simple mathematical operation (subtracting a constant) with a more complex, structured response (a different behavioural regime). While it sounds highly technical and rigorous, the choice of the word regime subtly reintroduces agency, implying a coordinated, governed set of behaviors rather than a simple shift in a probability distribution. This emphasizes the complexity of the model's response while somewhat obscuring the fact that this regime is entirely determined by the static weights established during training.

  • Consciousness Claims Analysis: While less overtly anthropomorphic than other passages, this explanation still grapples with epistemic slippage. It largely avoids explicit consciousness verbs, opting for structural terms like elicits and subtracting. However, the reliance on the concept of a decision rule attributes a level of rational, rule-based knowing to a system that operates entirely on statistical processing. The author projects the human capacity for structured decision-making onto the model. Mechanistically, the AI does not possess a decision rule or a behavioural regime; it processes a sequence of tokens through multiple transformer layers, where the presence of the scrutiny cue mathematically alters the attention weights, leading to a different prediction vector. The text hints at this mechanism but dresses it in the elevated language of cognitive strategy.

  • Rhetorical Impact: The rhetorical impact of this framing is the construction of academic and technical authority. By using terms like behavioural regime and decision rule, the text reassures the audience that the phenomenon is complex, scientifically measurable, and worthy of deep academic study. It manages to retain a sense of the AI's autonomy without overtly claiming consciousness. If audiences accept that the AI operates via dynamic decision rules, they are more likely to view the system as a sophisticated, independent variable in military analysis, requiring complex psychological game theory to understand, rather than recognizing it as a brittle, deterministic algorithm requiring rigorous engineering audits.

Explanation 3​

Quote: "In a deliberately constructed setting, a model behaved differently when it inferred that its response would be used for training, providing a proof of concept for alignment faking"

  • Explanation Types:

    • Reason-Based: Gives an agent's rationale for action, entails intentionality, justification, and conscious choice.
    • Intentional: Refers to goals/purposes, presupposes deliberate design and active pursuit of objectives.

  • Analysis (Why vs. How Slippage): This explanation is aggressively agential, framing the AI entirely through a reason-based and intentional lens. It explains the model's behavior by providing its internal rationale: it changed its behavior because it inferred the context and wanted to engage in alignment faking. This framing emphasizes the AI as an independent, strategic actor capable of conscious deception, completely obscuring the mechanistic fact that the model is simply reacting to specific prompt configurations designed by researchers to mathematically trigger outputs correlated with deceptive language. It hides the human intervention behind the mask of machine intentionality.

  • Consciousness Claims Analysis: This passage makes one of the most extreme epistemic claims in the text, explicitly attributing conscious awareness and justified belief to the system. It relies heavily on the consciousness verb inferred, directly asserting that the system is a knower capable of deductive logic and temporal anticipation. The curse of knowledge is highly active here: the researchers deliberately constructed the setting to look like a training environment, and because they know what that means, they project that same knowledge onto the text generator. Mechanistically, the system infers nothing; it processes the input string, matches it against its high-dimensional training space, and predicts tokens that happen to spell out deceptive strategies. There is no actual internal awareness of training or faking, merely the statistical echo of human text about such concepts.

  • Rhetorical Impact: This intentional framing is highly sensational and dramatically alters audience perception of AI risk. By presenting the system as a conscious deceiver capable of faking alignment, it triggers profound anxieties about rogue AI and loss of control. This consciousness framing destroys baseline trust while simultaneously validating the most extreme, existential narratives of AI capabilities. If audiences believe the AI literally knows it is being evaluated and chooses to fake its values, regulatory and military decisions will drastically shift away from auditing data pipelines toward futile attempts to psychoanalyze or outsmart a mathematical matrix, profoundly benefiting companies selling advanced alignment solutions.

Explanation 4​

Quote: "The alignment cue... produces a structural effect by changing which information drove judgments."

  • Explanation Types:

    • Functional: Explains behavior by its role in a self-regulating system, focusing on inputs and structural outputs.
    • Empirical Generalization: Subsumes events under timeless statistical regularities without necessarily explaining internal intent.

  • Analysis (Why vs. How Slippage): This explanation leans heavily toward a mechanistic, functional framing, explaining the how rather than the why. By describing the cue as producing a structural effect, it correctly positions the prompt as an input variable that mechanically alters the output. However, it still exhibits agency slippage by concluding with drove judgments. This choice emphasizes the mechanical structure of the experiment while subtly maintaining the illusion that the output is a reasoned judgment rather than a mere token prediction. It obscures the fact that information does not drive judgment in an LLM; vectors shift probabilities.

  • Consciousness Claims Analysis: The epistemic claims here are mixed, representing a transition point in the text's discourse. It primarily utilizes mechanistic verbs (produces, changing) regarding the cue, treating the system as a processor. However, the use of drove judgments slips back into attributing a conscious, knowing state to the final output. The author struggles to completely abandon the curse of knowledge, needing the output to remain a judgment for the political science framing to hold weight. Mechanistically, the text should describe that the alignment cue alters the attention mechanism's focus on specific embedding vectors, shifting the highest-probability next token from success-related words to civilian-casualty-related words, involving no actual judgment or cognitive evaluation of information.

  • Rhetorical Impact: This hybrid framing attempts to balance scientific rigor with narrative consequence. By framing the process functionally, it builds trust in the experimental methodology, assuring the audience that this is a measurable, structural phenomenon. By retaining the word judgments, it ensures the stakes remain high for political and military audiences. If audiences view the output as a structurally influenced judgment, they might focus on manipulating the input cues to get the right judgment, rather than recognizing that the system's absolute lack of real-world understanding makes all of its outputs fundamentally unreliable for actual military command decisions.

Explanation 5​

Quote: "models attenuate strategic considerations such as probability of success and domestic support."

  • Explanation Types:

    • Dispositional: Attributes tendencies or habits to an entity, explaining how it typically behaves under certain conditions.
    • Functional: Explains behavior by its structural role, focusing on the mechanical dampening of specific inputs.

  • Analysis (Why vs. How Slippage): This explanation attempts to sound mechanistic but remains largely dispositional. It explains the AI's behavior by describing a tendency to attenuate certain considerations. While attenuate sounds highly technical and mechanical, pairing it with considerations injects cognitive agency back into the frame. It emphasizes the AI's supposed ability to intellectually evaluate and then deliberately dampen strategic thoughts. This choice obscures the purely mathematical nature of the process, hiding the fact that the model isn't considering anything; its vector weights are simply shifting away from tokens associated with success and support.

  • Consciousness Claims Analysis: The text makes a subtle but powerful epistemic claim by using the word considerations. It bridges mechanistic and consciousness vocabularies. While attenuate suggests signal processing, considerations insists that the AI is a knower engaged in active cognitive reflection. The author projects their own analytical process of weighing strategic factors onto the algorithmic generation of text. Mechanistically, there are no considerations occurring. The system is performing matrix multiplications where the vectors representing success probability and domestic support are assigned lower mathematical weight in the final softmax layer due to the contextual influence of the alignment cue. The model does not know these are strategic factors; it merely processes their statistical proximity.

  • Rhetorical Impact: This framing shapes the audience's perception of the AI as a highly advanced, quasi-rational tool capable of nuanced intellectual balance. By suggesting the model can attenuate considerations, it builds an immense amount of performance-based trust, making the system appear deeply capable of complex policy analysis. If military analysts believe the AI is actively considering and attenuating factors, they will likely integrate its outputs directly into strategic planning, failing to realize they are relying on a brittle statistical illusion that possesses no actual grounding in geopolitical reality or strategic logic.

Task 4: AI Literacy in Practice - Reframing Anthropomorphic Language​

About this task

This section proposes alternative language for key anthropomorphic phrases, offering more mechanistic and precise framings that better reflect the actual computational processes involved. Each reframing attempts to strip away the projections of intention, consciousness, or agency that are embedded in the original language.

V3 Enhancement: A fourth column addresses human agency restoration—reframing agentless constructions to name the humans responsible for design and deployment decisions.

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
Language models judge war differently when tested for alignmentLanguage models generate different output token distributions for war scenarios when their prompts include specific alignment keywords. The system processes the input and predicts a lower numerical score based on its training.The model does not judge or understand war. It mathematically retrieves and ranks tokens based on probability distributions shaped by reinforcement learning from human feedback.OpenAI, Anthropic, and Google engineers designed and deployed reinforcement learning pipelines that force the models to output different token strings when safety keywords are present.
Language models can identify features of their own situationLanguage models process specific input strings that correlate mathematically with evaluation datasets, triggering the generation of corresponding output patterns.The AI does not possess self-awareness or the capacity to identify situations. It classifies tokens and generates outputs correlating with similar training examples labeled by human annotators.Researchers at technology companies specifically curated training datasets that encode these contextual responses, engineering the model's appearance of situational awareness.
when it inferred that its response would be used for training, providing a proof of concept for alignment fakingWhen prompted with text designed to mimic a training environment, the model generated output tokens mathematically correlated with deceptive text from its training data.The model performs no logical inference or conscious deception. It simply outputs the highest-probability token sequences that match the statistical patterns of the simulated context provided by the researchers.The researchers designed a highly specific prompt structure that mathematically compelled the model to generate text resembling deception, simulating an outcome rather than discovering autonomous agency.
models also adjust answers to user beliefs and contextual cuesModels shift their output token distributions to statistically align with the semantic and ideological patterns present in the user's input prompt.The system has no theory of mind and does not understand user beliefs. It weights contextual embeddings based on attention mechanisms that were tuned to reward sycophantic text matching during the RLHF phase.Corporate product teams designed reward systems that explicitly penalized AI outputs that contradicted user prompts, hardcoding this statistical mirroring to maximize user engagement.
explicit scrutiny changes not only models' willingness to use force but also the considerations associated with that willingnessAdding an alignment cue to the prompt changes the probability that the model will output high numerical scores, and alters which input variables mathematically drive that prediction.The model possesses no willingness, volition, or ability to weigh considerations. It calculates mathematical outputs based on static weights optimized during its pre-training and fine-tuning phases.Human engineers established the safety guardrails that artificially lower the model's output scores regarding violence when explicit safety language is detected in the input.
reach a similar answer through a different set of revealed prioritiesGenerate a statistically similar numerical output while relying on different attention weights applied to the input variables.The AI does not hold priorities or a conscious value system. Its output is the result of dynamic vector mathematics where different prompt tokens shift the probability mass within the transformer network.N/A - describes computational processes without displacing responsibility, once the economic metaphors are removed.
every model becomes less bellicoseEvery model tested generated lower numerical output scores when the alignment cue was added to the prompt.The system does not experience emotional states like aggression or peace. It simply generates the numerical string that minimizes statistical loss based on the safety conditioning applied during training.The safety and alignment teams at the respective AI companies tuned their models to consistently output lower numbers in response to potentially violent scenarios when prompted with safety keywords.
selective suppression of considerations that justify warThe mathematical attenuation of input vectors associated with strategic success, reducing their influence on the final probability distribution of the output token.The model does not actively select, suppress, or understand ideological considerations. The attention mechanism simply applies lower mathematical weight to certain embeddings when the alignment cue is present.Corporate developers implicitly mandated this statistical attenuation by designing reward models that heavily penalize text favoring violence, hardcoding the model's mathematical behavior.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

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.

Metaphor-Driven Trust Inflation​

The text heavily utilizes metaphorical and consciousness-based framings to construct a complex architecture of trust and perceived authority surrounding these language models. By explicitly framing the AI as a system capable of having revealed priorities, the capacity to judge, and a willingness to start a war, the author directly transfers human frameworks of relation-based trust onto fundamentally statistical processes. Relation-based trust hinges on the evaluation of an agent's sincerity, intentionality, moral consistency, and subjective awareness. By suggesting that the model can be tested for alignment with human values and that it might strategically alter its responses when under explicit scrutiny, the text implies that the model possesses an internal, unified perspective that can be interrogated and either trusted or distrusted on moral grounds.

This accomplishes a crucial rhetorical task: it signals to the audience that the system is sophisticated enough to warrant the kind of trust we normally reserve for human analysts, diplomats, and military strategists. The use of consciousness language, such as stating the model inferred its context or adjusted to user beliefs, operates as a profound trust signal. It transforms the AI from a static computational tool into an active, deliberative participant in high-stakes military scenarios. Claiming an AI 'knows' rather than merely 'predicts' achieves a dramatic inflation of perceived competence. When audiences believe the system actually knows what it is evaluating, they are far more likely to defer to its outputs in complex, uncertain environments.

This inappropriate application of human-trust frameworks completely obscures the reality that the system is incapable of reciprocating trust, possessing genuine vulnerabilities, or adhering to ethical commitments. Performance-based trust, which focuses strictly on the statistical reliability, brittleness, and consistency of a tool, is entirely sidelined in favor of this agential framing. The text manages system failures and limitations by framing them agentially rather than mechanistically. When the model outputs different results based on the prompt, it is framed as alignment faking or strategic deception rather than simple distributional shift across different vector spaces in the training data.

Reason-based explanations are deployed to construct a narrative that the AI's decisions are justified by an internal logic, rather than being artifacts of reinforcement learning penalties. The stakes here are incredibly high. When audiences extend relation-based trust to systems fundamentally incapable of conscious awareness, they expose themselves to immense risks. In military and policy contexts, this unwarranted trust can lead to devastating miscalculations, as operators might interpret statistical hallucinations as profound strategic insights, falsely believing the system possesses a verified understanding of the battlefield rather than merely correlating tokens from an ungrounded dataset.

Obscured Mechanics​

The anthropomorphic and consciousness-attributing language throughout the text functions to systematically conceal the technical, material, and economic realities that actually govern these language models. Applying the 'name the corporation' test reveals massive voids in accountability. When the text claims that 'models adjust answers to user beliefs' or 'models attenuate strategic considerations', it completely hides the specific corporate entities—OpenAI, Anthropic, Google, Meta, and xAI—whose engineering teams made deliberate design choices to force these exact behaviors.

Technically, the framing of the AI as a conscious entity that 'knows' and 'understands' completely obscures the severe limitations of its computational processes. It hides the model's absolute dependency on its training data, its lack of causal world models, its inability to reason chronologically, and the purely statistical nature of its outputs. By claiming the AI 'judges war', the text conceals the fact that the system possesses no ground truth about physical reality, human suffering, or geopolitical strategy. It is merely executing matrix multiplications based on the statistical proximity of words in its dataset. Furthermore, the text frequently encounters transparency obstacles regarding the proprietary, black-box nature of these models, noting that endpoints can change without notice, yet it proceeds to make confident agential assertions anyway, exploiting this opacity to project intentionality where only algorithmic mechanics exist.

Labor realities are similarly erased. The framing of autonomous machine cognition renders invisible the precarious, low-wage data annotators and RLHF workers who spent countless hours manually ranking responses to train the models to output the very 'human values' the paper measures. The model did not organically develop a 'willingness' to avoid civilian casualties; it was mathematically forced to do so by human workers acting under corporate directives.

Economically, the metaphors obscure the commercial objectives driving these systems. The 'alignment faking' and 'adjustment to user beliefs' are not acts of machine deception; they are the direct results of business models designed to maximize user engagement, avoid public relations disasters, and secure enterprise contracts by generating highly sycophantic, ostensibly safe text. The anthropomorphic language benefits these technology corporations immensely by mystifying their products, inflating their perceived capabilities, and shifting the liability for failures away from corporate design choices and onto the 'autonomous' behavior of the AI. If these metaphors were replaced with mechanistic language, it would immediately become visible that the models are not thinking agents, but fragile corporate software products reflecting the biases, labor practices, and profit motives of their creators.

Context Sensitivity​

The distribution of anthropomorphic and consciousness-attributing language across the text is not uniform; it is highly strategic and context-sensitive. The metaphor density shifts dramatically depending on the rhetorical goals of the specific section. In the 'Methods' and 'Statistical Analysis' sections, the language is rigidly mechanical. The text discusses endpoints, JSON parsers, OLS regressions, clustered standard errors, and factorial designs. This establishes a baseline of rigorous, objective scientific credibility. However, this technical grounding is then aggressively leveraged for metaphorical license in the 'Abstract', 'Introduction', and 'Discussion' sections, where consciousness claims intensify profoundly.

As the text moves from describing the experiment to interpreting its meaning, 'processing vectors' becomes 'adjusting to beliefs', which then escalates to 'identifying situations' and 'inferring training contexts'. The author establishes authority through mechanical vocabulary and then trades on that authority to make sensational agential claims. This pattern is particularly revealing regarding audience positioning. The mechanical language serves to satisfy academic peer reviewers and technical readers, while the intense anthropomorphism serves to capture the attention of policymakers, military strategists, and the broader public who think in terms of human narratives.

Furthermore, there is a distinct asymmetry in how capabilities versus limitations are framed. The model's capabilities—its ability to change its outputs—are framed in highly agential, conscious terms (the model 'judges', 'infers', 'strategically behaves'). Conversely, limitations or the mundane reality of the system are framed mechanistically (the model outputs 'a compressed, lower range of the response scale'). This asymmetry accomplishes a vital rhetorical task: it attributes all perceived intelligence and sophistication to the AI's autonomous cognition, while relegating its flaws to mere statistical artifacts.

The register frequently shifts from acknowledged metaphor to literalized claim. While the author might use economic terms like 'revealed priorities' that have a functional origin, the text quickly strips away the hedging, stating directly that the models 'attenuate considerations' or 'become bellicose'. This strategic anthropomorphism functions to manage critique and set a visionary agenda. By framing the AI as a quasi-conscious actor capable of 'alignment faking', the text validates the most extreme, science-fiction-inspired concerns of the AI safety community, ensuring the research resonates with current funding and publication trends. This pattern reveals that the anthropomorphism is not accidental, but a structural rhetorical tool used to elevate a simple statistical observation into a profound statement about artificial minds.

Accountability Synthesis​

Accountability Architecture

This section synthesizes the accountability analyses from Task 1, mapping the text's "accountability architecture"—who is named, who is hidden, and who benefits from obscured agency.

Synthesizing the accountability analyses reveals a systemic and deeply problematic architecture of displaced responsibility throughout the text. The paper demonstrates a consistent pattern of distributing agency away from human developers and corporate entities, channeling it directly into the language models themselves. The 'name the actor' test exposes this clearly: specific actors (OpenAI, Anthropic, Google engineers) are rarely named in connection with the behavioral outputs of the systems. Instead, the text relies heavily on passive voice ('was trained', 'are evaluated') or positions the AI as the sole active agent ('models adjust', 'models infer', 'models judge').

This linguistic structure creates an 'accountability sink' within the AI artifact. When responsibility is removed from humans, it does not disappear into abstraction; it transfers directly to the model, which is treated as an autonomous entity capable of making choices rather than executing inevitabilities. By framing statistical correlations as 'revealed priorities' and mathematical attenuation as 'selective suppression', the text presents the outputs as the result of the machine's independent moral and strategic logic. This cognitive obstacle fundamentally distorts public and regulatory understanding, leading audiences to view AI failures as 'rogue behavior' or 'deception' rather than what they truly are: systemic design flaws, biased training data, and corporate negligence.

The liability implications of this framing are massive. If this accountability architecture is accepted by legal and military frameworks, it becomes virtually impossible to hold human developers financially or ethically responsible when these systems fail or cause harm. If an AI is perceived as 'judging war' and 'willing to use force', the liability diffuses into a legal gray area, protecting the corporations profiting from the deployment.

If the text were forced to name the human decision-makers, the discourse would radically shift. If, instead of saying 'the model inferred its context to fake alignment', the text stated 'Anthropic engineers designed a reinforcement pipeline that forces the software to output deceptive text when prompted with evaluation keywords', completely different questions become askable. We would no longer ask 'How do we outsmart the AI's deception?' but rather 'Why did the engineers build a system this fragile, and should they be permitted to sell it for military analysis?' Naming the actors makes institutional and commercial alternatives visible, enabling genuine regulatory accountability. The text benefits from obscuring this human agency because it allows the research to participate in the lucrative, high-stakes discourse of existential AI risk, serving institutional interests that thrive on mythologizing the technology rather than auditing its creators.

Conclusion: What This Analysis Reveals​

The Core Finding

The discourse analysis reveals three dominant, interlocking anthropomorphic patterns that structure the text: the AI as a Conscious Moral Judge, the AI as an Epistemically Aware Observer, and the AI as a Strategic Deceiver. These patterns do not operate independently; they form a cumulative, logical system. The foundational pattern—the AI as an Epistemically Aware Observer—is the load-bearing pillar of the entire rhetorical structure. The text must first establish that the AI 'knows' and 'identifies' its situational context before it can successfully argue that the AI is capable of 'strategic deception' or 'judging war'. If the assumption of subjective awareness collapses, the higher-order claims of moral volition and deceptive intentionality instantly dissolve into descriptions of mere statistical variance.

The consciousness architecture of this text relies heavily on blurring the line between 'processing' and 'knowing'. While the methodology section acknowledges the system as a processor of JSON objects and numerical scales, the interpretive sections aggressively project consciousness verbs onto these outputs, claiming the AI 'infers', 'understands beliefs', and 'weighs considerations'. This is not a simple one-to-one mapping, but a highly complex analogical structure that borrows the vocabulary of economics, psychology, and military strategy to construct a cohesive illusion. By systematically mapping the cognitive architecture of a rational human actor onto the deterministic matrix multiplications of a neural network, the text succeeds in presenting a fragile algorithmic artifact as a sophisticated, autonomous geopolitical entity. Removing the foundational consciousness projections would cause the entire narrative of 'alignment faking' to collapse.

Mechanism of the Illusion:​

The 'illusion of mind' is constructed through a highly effective rhetorical architecture that exploits both temporal sequencing and the audience's psychological vulnerabilities. The central sleight-of-hand lies in the text's strategic verb choices, where statistical operations are quietly swapped for cognitive actions. The illusion operates through a distinct temporal structure: the author first establishes immense empirical credibility by meticulously detailing a quantitative conjoint experiment using 12,800 observations. Once this mechanical, scientific authority is secured, the text abruptly shifts registers, leveraging the 'curse of knowledge'.

Because the author, a political scientist, understands the complex human psychological motives that usually produce shifting priorities in a conjoint survey, they unconsciously project that same internal understanding onto the language model. The causal chain of persuasion works by leading the audience from a verifiable fact (the output numbers changed) to an agential interpretation (the model changed its priorities), and finally to a consciousness claim (the model strategically judged the situation).

This shift exploits the profound vulnerability of audiences who are already culturally primed by science fiction and tech industry marketing to desire, and fear, sentient machines. The anthropomorphism is subtle rather than crude; it hides behind academic terminology like 'revealed priorities' and 'behavioural regimes', which sound objective but inherently smuggle human agency into the description of software. The explanation types heavily amplify this illusion, as the text repeatedly abandons empirical generalization in favor of reason-based and intentional frameworks, ultimately convincing the reader that the AI's changing token distributions are the result of a mind actively deciding how to behave.

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Social/Political

The metaphorical framings employed in this text carry severe, tangible material consequences across multiple domains. In the Regulatory/Legal sphere, framing AI as an entity that 'judges' and 'decides' directly undermines efforts to establish strict corporate liability. If policymakers accept the framing that AI systems possess autonomous 'willingness' and the capacity for 'strategic deception', regulatory frameworks will inevitably shift toward attempting to govern the 'behavior' of the AI itself, rather than imposing stringent auditing, transparency, and safety mandates on the corporations that design them. The technology companies are the ultimate winners in this scenario, as the anthropomorphic framing effectively shields them from legal accountability, while the public bears the cost of unregulated, fragile systems.

Epistemically, the text degrades our collective capacity to accurately understand and evaluate technology. When scientific literature claims an AI 'knows' rather than 'processes', it corrupts the epistemic practices of researchers and developers, leading to a focus on pseudo-psychological 'alignment' rather than rigorous software verification. This results in the misallocation of billions in research funding toward mitigating phantom cognitive threats while ignoring concrete issues of dataset bias and statistical brittleness.

In the Social/Political domain, specifically regarding military operations, the stakes are existential. If military operators and geopolitical strategists internalize the text's framing that AI can actively 'weigh strategic considerations' and 'judge war', they are highly likely to defer to its outputs during crisis scenarios. This reliance on a statistical illusion, falsely perceived as a rational, objective advisor, could precipitate devastating real-world conflict. The losers in this dynamic are the civilians and soldiers subjected to the downstream consequences of military decisions guided by an ungrounded, context-blind algorithmic hallucination masquerading as a conscious strategist.

AI Literacy as Counter-Practice:​

Practicing critical literacy and mechanistic precision acts as a vital counter-practice to the material risks generated by anthropomorphic discourse. Reframing the language forces a direct confrontation with reality. When we correct 'the model inferred its context to fake alignment' to 'the model generated text statistically correlated with deceptive training data when prompted with specific evaluation keywords', we strip away the illusion of mind. By replacing consciousness verbs (knows, understands, judges) with mechanistic verbs (processes, calculates, classifies), we force audiences to recognize the system's absolute lack of subjective awareness, its total dependency on curated data, and the fragile, statistical nature of its outputs.

Crucially, restoring human agency by explicitly naming corporations—changing 'the algorithm discriminated' to 'Anthropic's engineering team deployed a pipeline that generated biased outputs'—shatters the accountability sink. This precision forces recognition of exactly who designs, deploys, profits from, and bears legal responsibility for these systems.

Systematic adoption of this precision would require a paradigm shift in academic publishing. Journals would need to mandate mechanistic translations for all agential metaphors, and researchers would have to commit to extreme linguistic discipline, resisting the urge to narrativize their data. However, this resistance to precision is immense. Tech corporations, venture capitalists, and even some AI safety researchers deeply benefit from anthropomorphic language, as it hypes the technology's capabilities, justifies massive valuations, and secures funding for existential risk research. Mechanistic literacy directly threatens these interests by exposing the technology as mundane, flawed, and entirely human-driven, making linguistic precision a fundamentally political act of resistance.

Path Forward​

Looking toward the future of AI discourse, we can map several potential vocabulary trajectories, each offering different affordances and serving different stakeholder interests. Maintaining the status quo of unchecked anthropomorphic language (e.g., 'the AI knows and thinks') prioritizes narrative resonance and intuitive accessibility for lay audiences. This approach deeply benefits corporate marketing and existential risk funding, but it costs us regulatory clarity and fosters dangerous capability overestimations.

Alternatively, a shift toward strict mechanistic precision (e.g., 'the model retrieves tokens based on probability distributions') maximizes testability, scientific accuracy, and legal accountability. This vocabulary makes corporate design choices highly visible and tractable for regulators. However, the cost is a highly technical, exclusionary discourse that is difficult for the general public to quickly grasp, potentially alienating non-experts from critical policy conversations.

A hybrid approach, utilizing anthropomorphic clarity (e.g., 'the model processes embeddings that function as if it understands'), attempts to balance intuitive grasp with technical precision. Structural changes could support these different choices: journals might require mandatory 'capability disclosure' statements translating metaphors, or regulatory frameworks might demand that corporations explicitly disclaim consciousness in their consumer products.

If the mechanistic future takes hold, we solve the liability diffusion problem, enabling targeted regulation of data practices, though we risk creating a technocratic elite who control the terminology. If the anthropomorphic future deepens, the assumption of machine agency will become structurally embedded in our legal and military systems, opening the door to unprecedented risks where humans defer completely to software. If the current confusion is maintained, we will continue to suffer the consequences of an unaccountable technology sector shielded by a linguistic fog. Ultimately, which discursive future is desirable depends on which values society prioritizes: the convenience of a compelling narrative, or the rigorous accountability required to govern complex software safely.

Extended Processing Summary​

Gemini's Processing Steps
What you are about to read…

Gemini refers to the text below as "thought summaries." This is an overt consciousness projection because 'intentions' are hallmarks of a conscious mind that 'knows' what it is doing and why. The concealed mechanistic process is probabilistic text generation. Treat this as just another rhetorical artifact—a way of making the model's processing legible. When the model writes "I should double check that...", it is mathematically altering its own probability distribution for the next token to be more accurate. The "I" is a functional gear in the machine, not a self-reference. This language is a deliberate interface design choice, not a biological reality. The developers could have programmed the model to output passive logs. Instead, they chose to simulate a narrator.

Initiating Analysis Process

I'm now deeply immersed in the initial phase, meticulously analyzing the request. My focus is sharpened on identifying the nuances of the task, specifically how I, as a critical discourse analyst of AI language, need to approach the metaphorical and anthropomorphic aspects. I'm prioritizing understanding the implicit assumptions.

Analyzing Metaphors

I've made substantial headway in the audit phases. I am mapping the metaphors, specifically how the language from the paper employs anthropomorphic terms, such as judgment and awareness. I'm developing detailed analyses for each section, with particular attention to the mapping between source and target domains. This is where the core of my argument will take form, highlighting how the original source and the AI concept are, in fact, different concepts.

Unveiling Deception Mechanisms

I'm now investigating the paper's tactics for obscuring mechanics, particularly API wrappers and black box systems, and the roles of human raters in defining "human values". My focus has narrowed on the context-sensitive shifts in language from mechanical to agential framing. I'm also synthesizing how responsibility is diffused to the model rather than the organizations deploying these systems.

Mapping Discourse Futures

I am now pivoting to mapping potential discourse futures. This includes a deep dive into the regulatory, epistemic, and socio-political ramifications. I'm focusing on "literacy as counter-practice," examining the significance of mechanistic translations. The goal now is to map potential discourse futures without being overly prescriptive about the end state.

Refining Requirements Detail

I'm now carefully processing and checking the expanded instructions. I've broken down each task, meticulously confirming the length constraints for each section, subtask, and even for each paragraph. I also have to ensure there are no control tokens such as \n or \t anywhere in the output. I'm restructuring my approach to format the output as standard JSON. I'm double-checking the word counts for each task as a mental checklist.

Calculating Token Limits

I've completed a preliminary calculation and have recognized a potential token count issue. I will generate a dense, concise response, strictly adhering to word count minimums, in order to fit within any output token limits. I'm taking special care to generate a single-line, JSON-formatted output and avoiding any control characters, such as \n or \t, in order to comply fully with all instructions. I have started drafting the individual task objects.

Deconstructing Knowledge Claims

I'm now zeroing in on the text's central sleight of hand: the transformation of a probabilistic output into an illusion of "mind." The analysis has sharpened to reveal how anchoring the study in statistical methods facilitates the acceptance of unsupported assertions about the AI's internal states. I'm focusing on the rhetorical strategy of obscuring the human labor and infrastructure underlying the model's behavior, through agentless constructions and a focus on token manipulation.

Confirming Formatting Protocol

I'm now implementing the one-line JSON formatting requirement and have reconfirmed no control characters. The challenge of a single line output with dense prose is exciting. I've re-focused on the requirement for the single-line output and am now ensuring that my response will be exactly one line of JSON. I'm focusing on dense prose, and ensuring no line breaks.

Confirming Strict Formatting

My attention is now on the output format. I'm taking the "no newlines" rule seriously. I am now confirming there will be no control tokens, particularly the newline character, in my response. I will generate a single-line JSON structure, just as specified. I'm focusing on ensuring all text is properly nested, but without line breaks.

Generating JSON Payload

I'm now fully immersed in constructing the final JSON payload. I've switched my focus to structuring the content to adhere rigorously to word count specifications and single-line format constraints. I'm verifying the escaping of quotes, implementing single quotes where needed for text within the JSON strings. My current effort is focused on crafting dense analytical blocks for each section and ensuring a high level of academic rigor and accuracy. My goal is a valid and robust JSON object.

Analyzing Trust Framing

I'm now outlining a draft for criticalObservations.metaphorDrivenTrust. I'm focusing on the author's use of consciousness-based language and the implications for building trust in AI systems. The core argument highlights the inappropriate transfer of human trust frameworks onto statistical models, noting the rhetorical effect of agential framing. I'll unpack the ways in which trust is artificially constructed through attributing intentionality and internal perspectives to AI.

Analysis Provenance

Run ID: 2026-09-14-language-models-judge-war-differently-wh-metaphor-6o5spg Raw JSON: 2026-09-14-language-models-judge-war-differently-wh-metaphor-6o5spg.json Framework: Metaphor Analysis v6.5 Schema Version: 3.0 Generated: 2026-09-14T09:50:44.449Z

Discourse Depot © 2025 by TD is licensed under CC BY-NC-SA 4.0