Skip to main content

SMMU: Benchmarking Social Intelligence of Multimodal Large Language Models

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.

Source Reading​

Genre: Peer-reviewed research · Author position: Academic or independent research

Argument: The text introduces SMMU, a benchmark designed to evaluate the "social intelligence" of Multimodal Large Language Models (MLLMs) over extended video sequences. The authors argue that existing benchmarks focus on short-horizon perceptual tasks and explicit cues, thereby failing to capture the complexity of real-world social understanding, which requires inferring implicit mental states, intentions, and beliefs over time. By proposing new joint evaluation metrics (Social Understanding Score and Social Agency Score), the text aims to penalize partial understanding and guess-based success, ultimately demonstrating that current frontier MLLMs still fall significantly short of human-level social intelligence and often rely on superficial post-hoc evidence matching rather than coherent reasoning.

Responsibility claims: The text holds the academic and developer communities responsible for advancing multimodal systems that can more accurately interpret complex social interactions. It directs its recommendations toward future researchers and developers, urging them to build models that "reason more coherently, update their beliefs as new evidence emerges, and ground their social interpretations." While the text does not explicitly assign blame for harms, it frames current MLLMs as active, albeit flawed, cognitive agents that exhibit "failures" to update understanding or engage in "premature anticipation." The text explicitly notes that models frequently fail to identify supporting evidence, locating the deficiency in the models' "social representations" and "reasoning" capabilities rather than directly in the training data curation or algorithmic design choices of their human developers.

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.

1. Model as Social Agent​

Quote: "do current MLLMs exhibit human-level social intelligence? i.e., Can MLLMs understand, reason and manage others in complex social interactions (Thorndike and Stein, 1937)?"

  • Frame: Model as socially capable human
  • Projection: This framing projects complex human cognitive and interpersonal capabilities onto statistical machine learning models. By asking whether models can "understand, reason and manage others," the text attributes epistemic and conscious states to computational processes. Understanding implies conscious comprehension of meaning, reasoning implies a deliberate logical process, and managing others implies an agentic, intentional capacity to influence human behavior. The text does not supply evidence for this level of agency; rather, it sets it up as a capability to be benchmarked. Mechanistically, at generation time, these models predict sequences of text tokens based on mathematical correlations learned during training from massive datasets of human language and video. They do not possess conscious awareness or the subjective experience required to actually "understand" a social interaction or "manage" a person.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text asks this question directly without hedging, presenting "understand, reason and manage" as literal capabilities that MLLMs might possess. I considered "Hedged/Qualified" since it is posed as a question, but the framing of the capabilities themselves lacks any metaphorical acknowledgment.)
  • Implications: This framing severely inflates the perceived sophistication and autonomy of MLLMs. By suggesting that AI could "manage others," it implies a level of agentic power and psychological insight that statistical systems simply do not possess. This can lead to unwarranted trust from users who might deploy these systems in sensitive interpersonal or clinical settings, believing the model possesses genuine social competence and empathy. Furthermore, it creates a liability ambiguity: if a model is thought to "manage" people, failures in that management might be attributed to the model's "poor social intelligence" rather than to the developers who inappropriately deployed a token-prediction engine for human management.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This framing obscures the human actors who design, train, and deploy these models. When asking if MLLMs can "manage others," the text hides the reality that companies like Google DeepMind, Alibaba, and Bytedance are the ones designing systems to output text that influences users. The agentless construction serves to position the AI as an independent actor, shifting focus away from the corporate decision to optimize models for human-like interaction. If actors were named, the question would shift to asking whether developers can train models to generate text that successfully mimics social management, restoring accountability to the creators. I considered "Partial" since "MLLMs" are named, but the actual responsible human entities are completely missing from this action.
Show more...

2. Model as Belief Updater​

Quote: "evaluate whether models correctly update their comprehension as events unfold and whether their reasoning and predictions remain consistent with their evolving social understanding."

  • Frame: Model as active epistemic agent
  • Projection: This metaphor projects dynamic human cognitive processes onto the static generation of text. The language of "updating comprehension," maintaining "consistent reasoning," and possessing an "evolving social understanding" attributes continuous epistemic and conscious states to the model. It suggests the model has an internal, subjective mental model of the world that it actively revises. Mechanistically, a deployed MLLM does not "update its understanding" in a conscious sense. When new evidence (tokens representing video frames or text) is added to its context window, the mathematical probabilities for generating subsequent tokens shift. The model processes new inputs statelessly or via attention mechanisms over the context window; it does not "understand" or hold "beliefs" that evolve.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The phrasing is presented as a literal description of what the benchmark evaluates. I considered "Hedged/Qualified" because the text later discusses "representations," but here the verbs "update," "comprehension," and "understanding" are used without any qualification.)
  • Implications: Attributing "evolving social understanding" to an AI system fosters the illusion of a continuous, conscious mind interacting with the user. This masks the brittle, stateless nature of most language model inference. When users believe a model "understands" them and "updates" its beliefs, they are more likely to over-trust its outputs, assuming its subsequent errors are minor misunderstandings rather than fundamental statistical hallucinations. This overestimation of capability can lead to dangerous reliance on AI in contexts requiring genuine situational awareness, while obscuring the fact that the system is merely generating statistically likely completions based on an updated prompt.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The agentless construction places the onus of "updating comprehension" entirely on the "models." This hides the engineering decisions made by developers at the named firms (e.g., Google DeepMind, Alibaba) regarding attention window size, context length limits, and training data biases. The text positions the model as the sole entity responsible for "consistent reasoning." Naming the actor would reveal that it is the system's architecture, chosen by engineers, that determines how well earlier tokens influence later token generation. I considered "Named" because the text evaluates specific models, but the human developers responsible for these mechanisms are entirely erased from the cognitive framing.

3. Model as Mentalizer​

Quote: "evaluating whether models can infer latent intentions, beliefs, perspectives, and relationships underlying human interactions"

  • Frame: Model as psychologist or theory-of-mind possessor
  • Projection: This framing projects the human psychological capacity of "theory of mind" onto a computational system. By claiming the model can "infer latent intentions, beliefs, perspectives," the text suggests the AI performs conscious perspective-taking and psychological analysis. The text explicitly grounds this in theories of human mentalizing. However, the evidence supplied only shows that models can select the correct multiple-choice text option correlated with a video clip. Mechanistically, generic to systems of this type, the model processes pixel data and transcripts, mapping them into a latent space where patterns associated with human labels (like "angry" or "intending to help") are mathematically clustered. The model predicts the most probable text string; it does not experience insight or infer subjective mental states.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The text directly states this as a capability to be evaluated, with no hedging. I considered "Explicitly Acknowledged" because the authors cite psychological theories of mentalizing, but they apply these theories to the model's capabilities literally rather than metaphorically.)
  • Implications: Using terms like "infer latent intentions" creates a powerful illusion of psychological depth. It suggests the model can "see into" human minds, which drastically inflates perceived sophistication. This framing is particularly risky because it encourages the deployment of AI for surveillance, automated hiring, and behavioral prediction, based on the false premise that the system genuinely comprehends human intent. In reality, the system is merely correlating surface-level audio-visual features with text labels, a process prone to severe bias and catastrophic failure when deployed outside its training distribution.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text attributes the action of inferring solely to the "models." This formulation hides the massive human labor of data annotators who originally labeled the "latent intentions" in the training data, as well as the developers who structured the loss functions. By making the AI the grammatical subject that "infers," the wording obscures the fact that the system is simply reproducing human annotations at scale. I considered "Partial" because the text mentions its own annotators for the benchmark, but the accountability for the models' underlying capabilities is entirely displaced onto the software itself.

4. Model as Flawed Reasoner​

Quote: "models frequently fail to identify the evidence supporting a social interpretation or to anticipate what will happen next, even when their comprehension is correct."

  • Frame: Model as student making logical errors
  • Projection: This metaphor projects the image of a human student who grasps a concept but fails to apply it correctly. The model is described as having "correct comprehension" but "failing to identify evidence" or "anticipate." This attributes both epistemic states (comprehension) and intentional states (anticipation, identifying) to the system. The text provides evidence of statistical failure on a benchmark, not of cognitive failure. Mechanistically, at generation time, the model's failure to select the correct downstream answer is a result of weak mathematical correlation in its learned parameters between the tokens of the first question and the tokens of the second question. The model lacks a centralized "comprehension" that it can consult; it simply computes conditional probabilities.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text uses "comprehension is correct" and "fail to identify" as literal descriptions of model behavior. I considered "Hedged/Qualified" since the text notes this might just be "post-hoc evidence matching" elsewhere, but in this specific quote, the cognitive terms are used without qualification.)
  • Implications: By framing statistical inconsistencies as a "failure to identify evidence" despite "correct comprehension," the text preserves the illusion that the model genuinely "comprehends" anything at all. It treats the model's output as a temporary lapse in logical reasoning rather than exposing the fundamental absence of understanding. This encourages users and policymakers to treat AI errors as "glitches" in an otherwise intelligent system, rather than recognizing that the system fundamentally lacks semantic grounding and will inherently produce inconsistent outputs when statistical patterns conflict.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The models are positioned as the actors who "fail to identify" and who possess "comprehension." This displaces accountability from the developers at firms like Google, Alibaba, and Bytedance who trained models that produce internally inconsistent outputs. The wording has the effect of making the AI the locus of failure, masking the reality that the human-designed architecture lacks mechanisms for enforcing logical consistency across generated tokens. I considered "Partial" due to the systemic critique, but the specific grammatical construction completely erases the human designers of these flawed systems.

5. Model as Hasty Judge​

Quote: "Premature Anticipation, in which models select the correct answer before the supporting evidence (plot twist) appears"

  • Frame: Model as impatient or biased observer
  • Projection: This framing projects the human trait of jumping to conclusions onto the model. "Premature Anticipation" suggests the model is actively trying to predict the future and making a cognitive error by acting too soon. It maps intentional goal-directed behavior onto statistical generation. Mechanistically, the model is not "anticipating"; when prompted with a truncated video and a multiple-choice question, it computes the highest probability token sequence based on its training distribution. If the "correct" answer is statistically common in the training data for similar contexts, the model will generate it regardless of whether the specific "plot twist" tokens are present in the current prompt. There is no impatience, only probability.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Hedged/Qualified (The text implicitly qualifies this by defining it as a specific error category ("in which models select the correct answer before..."), linking it to the mechanical action of selecting an answer. However, I considered "Direct" because the capitalized term "Premature Anticipation" strongly implies a cognitive state without explicit scare quotes.)
  • Implications: Labeling a statistical artifact "Premature Anticipation" anthropomorphizes a flaw in the training data distribution or the benchmark itself. It suggests the model is "trying" to get the right answer but just guessing too early. This masks the more troubling reality: the model's outputs are heavily driven by linguistic and cultural priors encoded in its weights, rather than by actual analysis of the visual evidence provided. Trust is inappropriately maintained because the error is framed as a relatable human mistake rather than a fundamental limitation of pattern-matching algorithms.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The action of "selecting" and "anticipating" is attributed solely to the models. This obscures the fact that the developers trained these systems on massive, uncurated datasets that instill strong priors, causing the model to output certain answers regardless of visual evidence. The agentless construction serves to blame the model for "anticipating" prematurely, hiding the structural design choices that prioritize statistical likelihood over grounded visual analysis. I considered "Named" because specific models are analyzed in the paper, but the human agency behind their behavior is missing from this description.

6. Model as Stubborn Believer​

Quote: "Failure to Update, in which they retain an outdated interpretation after new evidence (plot twist) is introduced."

  • Frame: Model as dogmatic or inflexible thinker
  • Projection: This metaphor projects the human cognitive flaw of stubbornness or confirmation bias onto the model. "Retain an outdated interpretation" and "Failure to Update" suggest an epistemic state where the model holds a belief and consciously refuses or fails to revise it when presented with new facts. Mechanistically, as generic to systems of this type, the model simply generates text based on the entire sequence of tokens in its context window. If earlier tokens strongly heavily weight the generation probabilities, or if the model's attention mechanism fails to adequately weight the newer "plot twist" tokens, it outputs text consistent with the beginning of the video. It does not "retain an interpretation"; it computes a conditional probability that favors earlier context.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text states "retain an outdated interpretation" directly as an explanation for model behavior. I considered "Hedged/Qualified" because it is presented alongside the mechanical "select the correct answer," but "interpretation" implies a subjective conscious state without any hedging.)
  • Implications: Describing statistical attention failures as a "failure to update" an "interpretation" reinforces the illusion of mind. It implies the model has the capacity for true interpretation but is just being inflexible. This impacts trust by making the system's errors seem like human cognitive biases rather than mechanical failures of self-attention across long context windows. It encourages users to try to "persuade" or "correct" the model as they would a person, wasting time and resources, while overestimating the system's baseline capacity for actual comprehension.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The models are framed as the subjects that "retain an outdated interpretation." This phrasing removes the human engineers who designed the attention mechanisms and context window processing capabilities that cause these specific failure modes. When the text says "they retain," it obscures the fact that companies like DeepMind and Alibaba deployed architectures that struggle to weigh late-context information against early-context priors. I considered "Partial" since it's a technical critique, but the grammar places full responsibility for the "failure" on the AI system.

7. Model as Epistemic Learner​

Quote: "We hope SMMU supports the development of multimodal systems that reason more coherently, update their beliefs as new evidence emerges, and ground their social interpretations"

  • Frame: Model as evolving conscious scholar
  • Projection: This framing projects a high-level epistemic state—holding and updating "beliefs"—onto the target domain of AI systems. It maps the human process of rational inquiry and belief revision onto the computational process of generating text that aligns with new prompt inputs. Mechanistically, AI systems do not have "beliefs." During training, their parameters are optimized to minimize loss on a dataset; during generation, they produce text based on those fixed parameters and the current context. A model cannot "update its beliefs" because it has no subjective awareness or justified truth claims; it can only shift its probability distribution for token prediction when new data is added to the prompt.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The phrase "update their beliefs" is used directly in the conclusion as a desired goal for future systems. I considered "Explicitly Acknowledged" if they had used quotes around "beliefs," but the presentation is entirely literal, treating beliefs as a property future systems should possess.)
  • Implications: This language sets a dangerously anthropomorphic goal for the field of AI research. By establishing "updating beliefs" as the target metric, it encourages the evaluation of models based on their ability to convincingly simulate human conviction and belief revision, rather than on their mechanistic reliability. This exacerbates the ELIZA effect, leading society to grant epistemic authority to systems that merely parrot the statistical shape of belief. It obscures the risk that these systems will generate highly convincing but factually ungrounded text, making them powerful tools for misinformation.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: This instance exhibits partial visibility because it explicitly calls for "the development of multimodal systems," which implies the existence of human developers who will do this developing. However, the resulting systems are immediately positioned as the active agents that will "reason," "update their beliefs," and "ground their interpretations." While the call to action addresses humans, the goal it sets focuses on transferring epistemic agency to the machine. I considered "Named" but no specific developers are identified, and "Hidden" was rejected because the word "development" inherently acknowledges human creators.

8. Model as Deceptive Plausible-Sounder​

Quote: "This suggests that current models may select the most superficially plausible answer rather than grounding their interpretations in coherent, evidence-based reasoning."

  • Frame: Model as corner-cutting student
  • Projection: This metaphor projects the human behavior of guessing or bluffing onto the model. It contrasts "selecting the most superficially plausible answer" with "grounding their interpretations in coherent reasoning," implying the model has a choice between a lazy heuristic and deep thought. Mechanistically, "selecting the plausible answer" and "grounding in reasoning" are the exact same process for an LLM: calculating the highest probability token sequence. The model does not choose to be superficial; it always outputs what is statistically probable given its training. The appearance of "plausibility without reasoning" is just a mismatch between human linguistic expectations and the model's correlation weights.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Hedged/Qualified (The text uses "suggests that" and "may select," which qualifies the claim about the model's behavior. I considered "Direct" because the verbs themselves (select, grounding) are presented without quotes, but the modal "may" provides a layer of epistemic hedging about the conclusion.)
  • Implications: Framing the model's output as a failure to engage in "coherent reasoning" implies that the model could engage in such reasoning if it "tried" harder or was built slightly better. This obscures the reality that LLMs fundamentally lack the architecture for grounded causal reasoning; they only perform pattern matching. This framing affects policy and trust by suggesting that scale or minor tweaks will unlock true reasoning, delaying the necessary implementation of strict safeguards against systems that are structurally incapable of knowing what they say.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text assigns the action of selecting plausible answers and failing to ground interpretations to "current models." This hides the reality that the models are functioning exactly as their human developers designed them to: by predicting highly probable (and thus often "plausible") text. The effect of this wording is to blame the artifact for being superficial, rather than critiquing the firms (e.g., OpenAI, DeepMind, Anthropic) whose reliance on next-token prediction inherently produces plausible rather than reasoned outputs. I considered "Partial" since "current models" refers to a known industry, but the specific grammatical agents are the models alone.

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: conscious human adult → Multimodal Large Language Model (application level)​

Quote: "do current MLLMs exhibit human-level social intelligence? i.e., Can MLLMs understand, reason and manage others in complex social interactions"

  • Source Domain: conscious human adult
  • Target Domain: Multimodal Large Language Model (application level)
  • Mapping: This mapping takes the relational structure of a socially competent human—someone who possesses conscious awareness, empathy, logical reasoning, and the intentional capacity to guide or influence others—and projects it onto an MLLM. It invites the assumption that the AI system possesses an internal psychological life, subjective experiences, and the ability to form deliberate strategies for interpersonal management. It positions the model not as a tool, but as an active, intelligent participant in a social dynamic.
  • What Is Concealed: This mapping profoundly conceals the statistical, mechanistic reality of token prediction and representation learning. It hides the fact that the system does not "understand" anything; it merely correlates pixels, audio waveforms, and text tokens with labels provided by human annotators. It obscures the immense proprietary training data and the RLHF (Reinforcement Learning from Human Feedback) labor required to make the model's outputs mimic social understanding. By framing processing as knowing, it creates a transparency obstacle, making it impossible for a lay reader to grasp that the system is a mathematical correlation engine rather than a mind.
Show more...

Mapping 2: active human learner → generation-time context processing​

Quote: "evaluate whether models correctly update their comprehension as events unfold"

  • Source Domain: active human learner
  • Target Domain: generation-time context processing
  • Mapping: The source domain is a conscious individual actively perceiving the world, integrating new information, and subjectively revising their internal understanding of a situation. Projected onto the target domain of an AI model processing a video, it invites the assumption that the model maintains a continuous, conscious "state of mind" that is actively modified as new frames are ingested. It implies a temporal continuity of subjective experience.
  • What Is Concealed: This conceals the stateless nature of model inference. Mechanistically, at generation time, a model does not "update" an internal subjective comprehension. It processes the entire context window (or uses attention caches) to compute probabilities for the next token. If new tokens (events) are added, the mathematical calculation changes, producing different output text. The metaphor hides the dependency on context window limits, attention mechanisms, and the statistical nature of the output, masking the fact that the process is entirely mathematical and devoid of conscious awareness.

Mapping 3: psychologist or empathetic human → computational pattern matching / classification​

Quote: "evaluating whether models can infer latent intentions, beliefs, perspectives"

  • Source Domain: psychologist or empathetic human
  • Target Domain: computational pattern matching / classification
  • Mapping: This maps the human capacity for "Theory of Mind"—the ability to deduce what another person is thinking or feeling based on empathy and shared human experience—onto the AI's classification architecture. It invites the assumption that the model can "look behind" the physical pixels and audio to perceive invisible mental states, suggesting the model has a deep, almost magical insight into human psychology.
  • What Is Concealed: It conceals the reality that the model is only "inferring" based on surface-level statistical correlations present in its training data. The model does not know what an "intention" is; it only knows that certain visual patterns (e.g., a person reaching for a cup) are statistically associated with the text token "intends to drink." This hides the dependency on human annotators who labeled the training data with these intentions. It obscures the lack of causal models in the AI, hiding the fact that the system is entirely dependent on the biases and subjective interpretations embedded in its proprietary training sets.

Mapping 4: student failing a reading comprehension test → statistical mismatch between prompt and generation​

Quote: "models frequently fail to identify the evidence supporting a social interpretation"

  • Source Domain: student failing a reading comprehension test
  • Target Domain: statistical mismatch between prompt and generation
  • Mapping: The source structure involves a conscious agent who grasps a general concept but lacks the logical rigor or attention to detail required to locate the specific facts that support their conclusion. Projected onto the model, it assumes the model possesses a unified "social interpretation" (a belief) but suffers a cognitive lapse in connecting it to "evidence." It maps human logical fallibility onto statistical generation.
  • What Is Concealed: This mapping hides the fact that the model does not have a unified "interpretation" independent of its token generation. Mechanistically, generating the correct answer to question A and the incorrect answer to question B does not mean it "failed to identify evidence" for A; it simply means the conditional probabilities for B did not align with A. It conceals the absence of a symbolic reasoning engine or a grounded causal model within the transformer architecture, obscuring the reality that these are parallel statistical outputs rather than a coherent chain of conscious thought.

Mapping 5: impatient or biased human observer → prioritization of training data priors over prompt context​

Quote: "Premature Anticipation, in which models select the correct answer before the supporting evidence (plot twist) appears"

  • Source Domain: impatient or biased human observer
  • Target Domain: prioritization of training data priors over prompt context
  • Mapping: This maps the human psychological state of impatience, jumping to conclusions, or confirmation bias onto the model's behavior. It invites the assumption that the model is actively trying to predict the outcome and makes a deliberate, albeit hasty, choice based on a "hunch." It frames a statistical output as an intentional, goal-directed mistake.
  • What Is Concealed: This metaphor conceals the overpowering influence of the model's training data distribution. The model isn't being "premature"; it is simply computing the highest probability output based on its weights. If the "correct" answer is a highly common trope or statistical pattern in the training data, the model will generate it even without specific prompt context. The anthropomorphism hides the mechanical reality of how pre-training biases (which are heavily guarded proprietary secrets) can override specific, situated visual evidence during generation.

Mapping 6: stubborn or dogmatic thinker → failure of attention mechanisms over long context​

Quote: "Failure to Update, in which they retain an outdated interpretation after new evidence"

  • Source Domain: stubborn or dogmatic thinker
  • Target Domain: failure of attention mechanisms over long context
  • Mapping: The source domain involves a person who holds a belief and refuses to change their mind when presented with new facts, implying a conscious choice to ignore evidence. Mapped onto the AI, it suggests the model has an established "interpretation" that it stubbornly defends or passively fails to revise. It implies the model possesses a continuous epistemic state.
  • What Is Concealed: This mapping conceals the technical limitations of attention mechanisms in transformer models when dealing with long contexts (like video). The model doesn't "retain an interpretation"; rather, its attention weights may fail to sufficiently prioritize the newer "plot twist" tokens compared to the massive accumulation of earlier context tokens. It hides the mechanical realities of context window degradation, "lost in the middle" phenomena, and the purely statistical nature of how sequence probability is calculated, replacing a technical flaw with a character flaw.

Mapping 7: ideal rational agent / scientist → future AI generation capabilities​

Quote: "update their beliefs as new evidence emerges, and ground their social interpretations"

  • Source Domain: ideal rational agent / scientist
  • Target Domain: future AI generation capabilities
  • Mapping: This maps the highest aspirations of human epistemology—rational belief revision, logical grounding, and careful interpretation—onto future AI models. It invites the reader to imagine a future system that operates as a conscious, objective truth-seeker, methodically adjusting its worldview based on empirical data.
  • What Is Concealed: It completely conceals the fundamental limits of the current deep learning paradigm. By framing the goal as "updating beliefs," it hides the fact that statistical models do not possess the architecture for belief, justification, or truth. It obscures the massive material and labor costs required to simulate such "grounding" through endless reinforcement learning and data curation. It also hides the commercial incentives of AI developers, who benefit from marketing their pattern-matching engines as objective, belief-holding entities, thereby obscuring the subjective, human-curated nature of the "truth" these models generate.

Mapping 8: lazy or deceptive student → token prediction optimizing for likelihood rather than logic​

Quote: "current models may select the most superficially plausible answer rather than grounding their interpretations in coherent, evidence-based reasoning."

  • Source Domain: lazy or deceptive student
  • Target Domain: token prediction optimizing for likelihood rather than logic
  • Mapping: The source structure is a human who has the capacity for deep, logical reasoning but chooses to rely on superficial heuristics or plausible-sounding bluffs to get by. Projected onto the model, it suggests the system is capable of "evidence-based reasoning" but is currently opting for a "superficial" shortcut. It maps moral or intentional failure onto a mathematical function.
  • What Is Concealed: This conceals the reality that generating the "most superficially plausible answer" is exactly and only what a language model is designed to do: it predicts the most statistically likely sequence of tokens. The mapping hides the fact that the model lacks the structural capacity to "choose" reasoning over plausibility. By attributing the failure to the model being "superficial," it obscures the responsibility of the developers who sell these systems as reasoning engines despite knowing they are fundamentally built only for plausibility.

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: "The reasoning question is conditioned on the ground-truth answer to the corresponding comprehension question."

  • Explanation Types:

    • Functional (Primary): Explains behavior by role in self-regulating system with feedback
    • Theoretical (Secondary): Embeds in deductive framework, may invoke unobservable mechanisms

  • Analysis (Why vs. How Slippage): This explanation frames the AI evaluation process mechanistically (how it is structured). It describes the architectural design of the benchmark itself, explaining that the evaluation of "reasoning" is structurally dependent on the successful completion of the "comprehension" task. By using terms like "conditioned on" and "ground-truth answer," the text emphasizes the formal, logical dependencies built into the evaluation framework. This choice emphasizes the rigor and systematic nature of the benchmark design. It obscures any agential qualities of the model in this specific sentence, treating the model's outputs merely as data points that pass through a predefined conditional logic gate.

  • Consciousness Claims Analysis: This specific passage does not attribute conscious states to the AI system; rather, it describes the structure of the evaluation questions. The verbs present ("is conditioned") are mechanistic and apply to the benchmark, not the model. There is no knowing versus processing conflation here, placing this strictly at the "Computational activity" level of the agency gradient. Mechanistically, generic to systems of this type, when the benchmark evaluates the model, it provides the text of the comprehension question and its correct answer as part of the prompt context before asking the reasoning question, thereby constraining the generation probabilities for the subsequent tokens.

  • Rhetorical Impact: This mechanistic framing shapes the audience's perception of the benchmark as highly scientific, rigorous, and objective. By using formal language ("conditioned on"), it reassures the reader that the evaluation is tightly controlled. This builds trust in the benchmark's methodology, contrasting with the more anthropomorphic language used to describe the models' failures, thereby positioning the human researchers as authoritative and precise.

Show more...

Explanation 2​

Quote: "This suggests that current models may select the most superficially plausible answer rather than grounding their interpretations in coherent, evidence-based reasoning."

  • Explanation Types:

    • Dispositional (Primary): Attributes tendencies or habits
    • Intentional (Secondary): Refers to goals/purposes, presupposes deliberate design

  • Analysis (Why vs. How Slippage): This explanation frames the AI agentially (why it acts). It attributes a disposition or tendency to the models—preferring "superficially plausible" answers over "coherent reasoning." This choice emphasizes the behavioral output of the system in terms of human cognitive flaws (laziness, superficiality). It obscures the mechanistic reality that selecting the "plausible" answer is precisely what the model's loss function optimized it to do during training (predict the most likely tokens). It frames a mathematical inevitability as a behavioral choice or tendency.

  • Consciousness Claims Analysis: The passage heavily attributes conscious states. It contrasts "select" (mechanistic/delegated) with "grounding their interpretations in coherent... reasoning" (consciousness/epistemic verbs). It projects knowing onto processing, locating the system at the "Intentional-state attribution" level on the agency gradient. The author projects human cognitive frameworks onto the system (curse of knowledge), assuming that a lack of logical consistency implies a choice to be superficial. Mechanistically, generic to models of this type, the model generates text token by token based on probabilities shaped by its training data. If a sequence of tokens is statistically highly probable (plausible), it is generated. The model has no internal mechanism for "coherent reasoning" other than attending to previously generated tokens.

  • Rhetorical Impact: This framing severely impacts audience perception by maintaining the illusion of agency even in failure. By suggesting the model "selects the superficially plausible answer" instead of "reasoning," it implies the model is capable of reasoning but is currently underperforming. This affects trust by making audiences believe the system is highly sophisticated but occasionally "lazy," leading them to rely on it in high-stakes situations while mistakenly assuming that a better prompt or a future update will force the model to "try harder" and actually reason.

Explanation 3​

Quote: "Gemini-3.1-Pro exhibits the strongest prediction dependency, with correct comprehension increasing prediction accuracy by 38.1 percentage points."

  • Explanation Types:

    • Empirical Generalization (Primary): Subsumes events under timeless statistical regularities

  • Analysis (Why vs. How Slippage): This explanation frames the AI mechanistically (how it typically behaves). It uses empirical data to describe a statistical regularity in the performance of a specific model (Gemini-3.1-Pro). The choice of "prediction dependency" and quantitative metrics ("38.1 percentage points") emphasizes the measurable, statistical relationship between different outputs. This framing effectively highlights the computational reality of the model's behavior, treating it as an artifact whose performance can be quantified. It obscures any internal "reasoning" process, reducing the model's "comprehension" and "prediction" to correlated success rates.

  • Consciousness Claims Analysis: While the overall structure is empirical, the passage relies on anthropomorphic shorthand verbs: "comprehension" and "prediction." It assesses processing but labels it as knowing, locating it at "Goal-directed behavior" (successfully completing the task) but bordering on "Epistemic states" due to the vocabulary. Mechanistically, based on reasonable inference from the benchmark design, when Gemini-3.1-Pro is evaluated, generating the tokens that correspond to the correct "comprehension" answer changes the context window in a way that significantly shifts the mathematical probabilities toward generating the correct "prediction" answer. The model is not comprehending; its token outputs are statistically correlated.

  • Rhetorical Impact: This empirical framing temporarily strips away some of the anthropomorphic illusion, grounding the reader in the reality that model performance is a matter of statistical accuracy and dependency. It builds trust in the authors' rigorous evaluation methods. However, because it still uses words like "comprehension," it maintains a subtle rhetorical link to human cognition, allowing audiences to mistakenly believe that the 38.1% increase represents an improvement in the model's actual "understanding" of the scene.

Explanation 4​

Quote: "Most comprehension-transition errors arise from either Premature Anticipation, in which models select the correct answer before the supporting evidence (plot twist) appears..."

  • Explanation Types:

    • Dispositional (Primary): Attributes tendencies or habits
    • Intentional (Secondary): Refers to goals/purposes, presupposes deliberate design

  • Analysis (Why vs. How Slippage): This explanation blends mechanistic categorization with agential framing. By defining a category of error ("Premature Anticipation"), it appears theoretical, but the definition attributes a behavioral tendency (anticipating too early) to the models. This choice emphasizes the temporal mismatch between the model's output and the video's timeline, framing it as an active mistake by the model. It obscures the underlying statistical cause: that the model's training priors are overriding the specific visual context provided in the prompt.

  • Consciousness Claims Analysis: The passage attributes intentional states through the term "Anticipation." It locates the system at "Intentional-state attribution" on the agency gradient. The author projects human test-taking behaviors onto statistical generation. Mechanistically, generic to systems of this type, when the model generates text in response to the truncated video, it computes probabilities based on its parameters. If the "correct" answer (which occurs after the plot twist) is highly correlated with the setup of the video in the training data, the model will generate it immediately, regardless of the missing "evidence" tokens in the prompt. It is not "anticipating"; it is calculating unconditional probabilities that overpower the conditional evidence.

  • Rhetorical Impact: Framing statistical artifacts as "Premature Anticipation" shapes audience perception by anthropomorphizing model failure. It suggests the model is "smart" but "hasty." This affects trust by minimizing the severity of the error—treating it as a relatable human flaw rather than a fundamental inability to ground outputs in presented evidence. It leads audiences to overestimate model autonomy and capability, believing the model "knew" what was going to happen, rather than recognizing it merely regurgitated a statistical trope.

Explanation 5​

Quote: "These results suggest that models' representations of the social scene are only weakly coupled with how they answer different questions about the same video, consistent with reliance on post-hoc evidence matching..."

  • Explanation Types:

    • Theoretical (Primary): Embeds in deductive framework, may invoke unobservable mechanisms
    • Empirical Generalization (Secondary): Subsumes events under timeless statistical regularities

  • Analysis (Why vs. How Slippage): This explanation frames the AI largely mechanistically (how it is structured). It deduces the internal architecture of the model's processing ("representations... are weakly coupled") based on empirical results. The choice to describe behavior as "post-hoc evidence matching" effectively demystifies the model's capabilities, emphasizing that it does not possess coherent, forward-looking reasoning. This framing obscures the agential illusion built up earlier in the text, offering a sobering, structural critique of how large language models actually function when generating responses to sequential questions.

  • Consciousness Claims Analysis: This passage actively resists attributing conscious states, using mechanistic terms like "weakly coupled" and "post-hoc evidence matching." It correctly locates the system at "Computational activity" on the agency gradient. It breaks the curse of knowledge by recognizing that correct outputs do not imply human-like reasoning. Mechanistically, generic to systems of this type, the model's hidden states ("representations") generated from the video do not form a centralized, consistent knowledge base. When prompted with different questions, the attention mechanism computes different probabilities for the output tokens, leading to inconsistent answers because the model is matching patterns locally for each prompt rather than consulting a unified "understanding."

  • Rhetorical Impact: This theoretical and mechanistic framing significantly shifts audience perception away from anthropomorphism. By explicitly naming "post-hoc evidence matching," it diminishes the perceived autonomy and reliability of the AI. It warns users that the system cannot be trusted to maintain coherent logic across interactions. If audiences internalize that AI merely "matches evidence post-hoc" rather than "knows," their trust in the system for critical, multi-step decision-making tasks would rightly decrease, shifting reliance back to human oversight.

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.

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
Can MLLMs understand, reason and manage others in complex social interactionsCan MLLM architectures process multimodal inputs to generate text that aligns with human labels for complex social interactions, and produce outputs that users might rely on to manage interpersonal situations?Mechanistically, generic to systems of this type, models process pixel data, audio, and text transcripts through transformer layers to predict subsequent tokens. They do not possess conscious awareness to "understand" or "reason," nor do they have intentionality to "manage others."

Basis: Generic to system type		Can developers at firms like Google and Alibaba train MLLMs to generate text that aligns with human labels, and can human users successfully deploy these systems to assist in managing complex social interactions?
models must integrate relevant evidence across extended interactions to understand the videoThe system must process tokens spanning extended video interactions to generate classifications that match human evaluations.Mechanistically, generic to systems of this type, the model uses attention mechanisms to weight tokens from earlier in the video sequence when calculating probabilities for generating the current output token. It processes data statistically; it does not "understand" a video.

Basis: Generic to system type		—
evaluate whether models correctly update their comprehension as events unfoldevaluate whether the text generated by the model changes appropriately when new video frames are added to its context window.Mechanistically, generic to systems of this type, at generation time, adding new tokens representing unfolding events to the model's prompt alters the mathematical probabilities for subsequent token prediction. The model generates new text based on these probabilities; it does not possess "comprehension" to update.

Basis: Generic to system type		—
evaluating whether models can infer latent intentions, beliefs, perspectivesevaluating whether models generate text that correctly classifies human behaviors associated with latent intentions, beliefs, and perspectives based on patterns in their training data.Mechanistically, generic to systems of this type, models map multimodal input data into a latent space where it is mathematically correlated with text strings representing intentions or beliefs, based on human-annotated training data. The model predicts text; it does not "infer" subjective mental states.

Basis: Generic to system type		evaluating whether developers have successfully trained models to output text that correlates with the latent intentions, beliefs, and perspectives labeled by human annotators.
models frequently fail to identify the evidence supporting a social interpretationthe text generated by the models is frequently mathematically uncorrelated with the specific visual evidence in the prompt, even when the model generates a statistically common social label.Mechanistically, based on reasonable inference from the model's architecture, the self-attention weights prioritize heavily represented statistical priors from the training data over the specific tokens representing the visual evidence in the prompt, resulting in a generated output that lacks logical consistency.

Basis: Reasonable inference		the models built by developers frequently generate text that is mathematically uncorrelated with the specific visual evidence.
Premature Anticipation, in which models select the correct answer before the supporting evidence (plot twist) appearsPrior-dominated Generation, in which the model generates the target answer based on statistical priors in its training data before the specific tokens representing the plot twist are introduced to the prompt.Mechanistically, generic to systems of this type, the model calculates the probability of the target answer. If the context prior to the plot twist is strongly correlated with that answer in the training data, the model generates it immediately, rather than waiting or "anticipating."

Basis: Generic to system type		Prior-dominated Generation, in which the models designed by these firms generate the target answer based on the statistical priors embedded in their training datasets.
Failure to Update, in which they retain an outdated interpretation after new evidenceContext Neglect, in which the model's generated output continues to be heavily weighted by earlier tokens, failing to shift probabilities when new tokens are added.Mechanistically, generic to systems of this type, the attention mechanism fails to sufficiently weight the newly added tokens compared to the larger volume of preceding tokens, causing the model to generate text consistent with the beginning of the prompt. It does not hold an "interpretation."

Basis: Generic to system type		—
models that reason more coherently, update their beliefs as new evidence emergessystems that generate logically consistent sequences of text and shift their output probabilities appropriately when new tokens are introduced to the context.Mechanistically, generic to systems of this type, models generate text based on parameters and context; they do not have subjective "beliefs" to update or the capacity to "reason." Improvement means better mathematical alignment between prompt context and token generation.

Basis: Generic to system type		we hope developers build multimodal systems that generate logically consistent sequences of text and shift their output probabilities appropriately when new tokens are introduced.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

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.

Metaphor-Driven Trust Inflation​

The text constructs authority and influences perceived reliability through metaphors that project human cognitive frameworks onto statistical systems. By adopting the vocabulary of "social intelligence," "comprehension," and "reasoning," the framing encourages relation-based trust in systems that are only capable of performance-based reliability. When the text claims an AI can "infer latent intentions" or "manage others," it signals to the reader that the system possesses the sincerity, empathy, and psychological insight necessary for human-trust frameworks. This wording can have the effect of inappropriately transferring expectations of human interpersonal competence onto a correlation engine. The risks emerge sharply when the text discusses system failures. Instead of framing errors mechanistically (e.g., "the transformer's attention mechanism failed to weight recent tokens"), the text frames them agentially: the model "fails to identify evidence" or "retains an outdated interpretation." These reason-based and intentional explanations construct a sense that the AI's decisions are part of a justified, conscious thought process that merely hit a snag. If audiences extend relation-based trust to these systems, believing they truly "comprehend" but occasionally suffer from "premature anticipation," they may deploy them in sensitive contexts like healthcare or automated management, assuming the model can "update its beliefs" when corrected. The text ultimately warns that models are "weakly coupled" and rely on "post-hoc evidence matching," but the dominant consciousness framing makes it difficult for readers to fully integrate this warning, preserving an unearned sense of authority.

Obscured Mechanics​

This framework holds that commercial incentives, data annotation and human feedback labor, environmental and compute costs, and developers' limited disclosure of training data, objectives, and feedback processes are standing conditions of the technology. The anthropomorphic and consciousness-attributing language in the text systematically conceals these realities. When the text says "models can infer latent intentions" or "models... select the correct answer," it obscures the massive, invisible labor of the data annotators who originally mapped visual cues to text labels like "intending to help." The text names the student annotators who built the benchmark, but the proprietary RLHF labor required to train models like Gemini (Google DeepMind), Qwen (Alibaba), and Kimi (Moonshot) remains entirely hidden behind the illusion that the "model" is doing the inferring. Technically, the claim that an AI "knows" or "understands" hides the reality that these systems rely entirely on statistical token prediction within bounded context windows; they lack ground truth and causal models. The framing of "Premature Anticipation" obscures the fact that pre-training data (which developers keep proprietary) instills overpowering statistical priors that cause the model to ignore new prompt context. Economically, framing the systems as possessing "social intelligence" benefits the commercial objectives of the named firms, masking their profit motives behind a veneer of scientific pursuit of artificial cognition. If these metaphors were replaced with mechanistic language—stating that "Alibaba's Qwen generates text statistically correlated with human-labeled intentions"—the dependency on human labor, the brittleness of the statistical architecture, and the corporate ownership of the data would become immediately visible.

Context Sensitivity​

The distribution of anthropomorphic and consciousness-attributing language is highly asymmetrical across the text, fundamentally shaping the reader's picture of the system. In the abstract, introduction, and conclusion—the sections dedicated to vision-setting and defining the benchmark's scope—consciousness claims intensify dramatically. Here, "processes" becomes "understands," which in turn becomes "knows" and "updates beliefs." This concentration of agential language in the framing sections establishes a default assumption of high-level cognitive capability. However, when the text moves into the technical methodology and error analysis, the language shifts. The capabilities are described in highly agential terms ("AI knows when to intervene," "infers intent"), while the limitations and underlying mechanics are suddenly described in mechanical terms (models rely on "post-hoc evidence matching" and exhibit "weakly coupled" representations). This asymmetry—capabilities in agential terms, limitations in mechanical terms—leaves the reader with a distorted picture: the system is viewed as an intelligent agent that occasionally suffers from mechanical glitches, rather than recognizing that the "intelligence" itself is merely a mechanical illusion. The register shifts noticeably: what begins as an acknowledged evaluation metric (Social Understanding Score) becomes literalized as the model's actual "evolving social understanding." This inherited register of the field positions the implied audience to accept the models as proto-conscious entities, while providing just enough technical jargon in the results section to maintain academic credibility without puncturing the overarching illusion.

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.

The text's vocabulary constructs an architecture of displaced responsibility that sits in subtle tension with its own explicit arguments. In its source reading, the text addresses human developers, urging the creation of systems that reason more coherently, thereby implicitly placing responsibility on the creators. However, the linguistic distribution of agency systematically diffuses this responsibility. Passive voice and agentless constructions concentrate heavily around failure modes: models "fail to update," "select the most superficially plausible answer," and exhibit "premature anticipation." By naming the models (Gemini, Qwen, Seed) but not the corporate entities (Google DeepMind, Alibaba, Bytedance) that designed their architectures and curated their training data, the text creates an "accountability sink." Responsibility transfers to the AI as an independent agent; it is the "model" that stubbornly "retains an outdated interpretation." This framework reads this displacement as serving the institutional and commercial interests of AI developers, as it naturalizes structural flaws in transformer architectures as individual "cognitive errors" made by the machine. If this framing is accepted, liability for harms caused by these systems in social contexts (e.g., misreading a user's intent) stops at the AI as a delegate. If human decision-makers were named—if the text stated that "Google DeepMind deployed an architecture that statistically overrides visual evidence with training priors"—it would make visible the alternatives (e.g., requiring causal modeling before deployment) and make structural accountability possible.

Conclusion: What This Analysis Reveals​

The Core Finding

Two dominant anthropomorphic patterns emerge in this text: the AI as a Conscious Social Agent (understanding, managing, intending) and the AI as an Epistemic Learner (updating beliefs, interpreting, failing to reason). These patterns are deeply interconnected; the foundational pattern is the Epistemic Learner. The assumption that the system possesses a continuous, subjective mental state (an "evolving social understanding") must be accepted for the projection of complex social agency to work. If the model does not "know" anything, it certainly cannot "manage others." The consciousness architecture of the text consistently conflates mathematical processing with conscious knowing, setting "updating beliefs" as the ultimate benchmark goal. The text's vocabulary is in tension with its own technical argument: while the authors rigorously demonstrate that models merely perform "post-hoc evidence matching" (a mechanical failure), their vocabulary insists on describing this as a failure to "ground interpretations in coherent reasoning" (a cognitive failure). This language complicates their call for better models by misdiagnosing a fundamental architectural limitation as a mere lapse in the machine's judgment.

Mechanism of the Illusion:​

This framework holds that computational systems process rather than know. The text produces the illusion of mind through a sophisticated, largely unacknowledged causal chain of metaphorical mappings. The central move occurs early, when the text defines "social intelligence"—a deeply human, relational capacity—as a capability to be benchmarked in LLMs. By doing so, the text activates the curse of knowledge: because humans use theory of mind to infer intentions from videos, the authors attribute that same psychological process TO the statistical system that outputs the correct text label. The illusion capitalizes on audience predispositions to anthropomorphize fluent language (the ELIZA effect). The temporal structure is crucial: the text first establishes the AI as a "knower" in the introduction, asking if it can "understand and manage others." Once the audience accepts this premise, the subsequent reason-based explanations of failure (e.g., "Premature Anticipation") land perfectly. The wording can have the effect of making the reader view statistical misalignments not as evidence that the machine is mindless, but as relatable human flaws—impatience or stubbornness—thereby cementing the illusion that there is a mind there to be flawed.

Material Stakes:​

Categories: Epistemic, Social/Political, Regulatory/Legal

The framing of AI as possessing "social intelligence" and the capacity to "update beliefs" carries immediate material consequences. Epistemically, if audiences believe AI "knows" rather than "processes," they will increasingly rely on these systems as objective truth-arbiters in complex social situations. This shift devalues human contextual judgment and relies on the unverified, proprietary statistical priors of models built by a few corporations. Socially, believing a model can "infer latent intentions" paves the way for the deployment of these systems in automated hiring, policing, and psychiatric evaluation. If a decision-maker believes the AI truly "understands" intent, they will delegate life-altering judgments to a system that is merely correlating pixels to text, disproportionately harming marginalized groups whose behaviors fall outside the model's training distribution. In the regulatory domain, the "accountability sink" created by blaming the "model" for "failing to update" protects developers from liability. It allows firms to market their systems as "socially intelligent" agents while avoiding the regulatory scrutiny that would accompany the plain admission that their products are brittle correlation engines prone to catastrophic context neglect. The developers win by maximizing perceived capability and minimizing liability, while the public bears the cost of unaccountable, automated social judgments.

AI Literacy as Counter-Practice:​

Practicing critical literacy against this discourse requires systematic human agency restoration and epistemic correction. By reframing "models update their beliefs" to "systems shift output probabilities based on context," we force the recognition that the AI has no awareness and is entirely dependent on its mathematical architecture. Reframing "models fail to identify evidence" to "the models built by developers generate text mathematically uncorrelated with the visual evidence" forces recognition of the human actors who design, deploy, and profit from these architectures. This precision directly counters the material stakes by making the limitations of the technology starkly visible, preventing the unwarranted delegation of social judgment to machines. However, implementing this precision faces severe institutional resistance. Marketing departments, funding narratives, and the disciplinary habits of computer science all incentivize anthropomorphic shorthand because it makes complex systems accessible and sounds groundbreaking. Precision is costly for AI developers because it demystifies their products, revealing them as limited tools rather than artificial minds. Anthropomorphic language serves the structural interests of the industry, regardless of individual authors' intent, by maintaining the hype cycle and obscuring the need for regulatory intervention.

Path Forward​

Looking ahead, the discourse ecology surrounding AI social evaluation faces divergent futures. The status quo—a hybrid of highly anthropomorphic capabilities and slightly mechanical limitations—serves the AI industry by maximizing hype while providing academic cover. It makes the rapid deployment of models tractable but renders accurate public risk assessment nearly impossible. Conversely, adopting rigorous mechanistic precision ("the system generates text statistically associated with human labels") makes the absence of mind visible and supports strict accountability, but at the cost of narrative resonance and intuitive accessibility for lay readers. A third approach, anthropomorphic clarity (using "thinks" but heavily caveating it as purely functional), attempts to balance precision with accessibility but often fails as audiences ignore the caveats and internalize the metaphor. Structural changes could support better choices: journals could require mechanistic translations of cognitive claims in abstracts, and regulatory frameworks could mandate transparency about the statistical nature of "inferences." If mechanistic precision becomes the norm, the problem of over-trust in AI social management could be solved, though it might slow adoption and funding. If anthropomorphic language deepens, we risk embedding the assumption that machines are our epistemic equals, making society highly vulnerable to automated deception and the unaccountable delegation of human social life to corporate algorithms. The choice of vocabulary fundamentally dictates whether we view ourselves as deploying mathematical tools or living alongside artificial minds.

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.

Defining the Focus

I'm now zeroing in on defining the precise parameters of this Critical Discourse Analysis. My focus is the AI metaphor and anthropomorphism within the "SMMU" paper. I'm considering what key questions I need to answer. I am defining how to best approach the study, in order to get the most insightful analysis.

Evaluating the Source

I've begun analyzing the source text. It's a peer-reviewed research paper from EMNLP 2026, focused on a new benchmark, SMMU, for evaluating social intelligence in MLLMs. I'm noting the authors' academic position and their argument for the need for extended interaction analysis. They're specifically assessing the capabilities of these models within a social context.

Analyzing the Framing

I'm now diving deeper into the framing. The paper challenges the AI community to create models that "reason coherently" and "update beliefs," holding current models accountable for "premature anticipation." I'm identifying several key instances of metaphor and anthropomorphism within the text, like framing models as needing to "understand" and "manage" in social contexts. I'm focusing on instances describing reasoning as a human quality.

Auditing Social Framing

I'm now auditing specific instances of anthropomorphism and framing within the paper, specifically within social contexts. I'm focusing on how the authors describe model behavior and reasoning. I'm noting specific phrases such as "fail to identify the evidence," "premature anticipation," and "update their beliefs," to reveal the paper's characterization of AI as having conscious, epistemic properties. I need to make sure I am distinct between the mechanistic processing and conscious knowing.

Mapping the Instances

I've mapped five instances within the paper to Brown's Typology. I've categorized them as: Functional/Theoretical, Dispositional/Intentional, Empirical Generalization, Functional/Dispositional, and a final instance. I'm noting each phrase's specific type within the framework.

Reframing the Focus

I'm now rewriting key instances to emphasize mechanistic realities. I'm replacing phrases that anthropomorphize models with descriptions of their actual processes. For example, "models update their beliefs" is now "the system adjusts its generated text based on new tokens in the input context window." I'm also mapping all instances within the proper enum type. I'm focusing on preserving human agency in each reframing.

I'm now tracing the agency slippage, and noting the move from mechanical descriptions to treating models as agents. I am starting to discuss how terms like "social intelligence" mask their statistical nature.

Analyzing the Accountability

I'm now zeroing in on the accountability aspect. The paper's framing obscures the mechanics, specifically by not naming the corporations behind the models. I'm noting the framework's focus on compute costs and annotator labor, but the training data and labor are hidden. I'm also examining how anthropomorphism increases when describing goals, like "infer latent intentions," but decreases when explaining failures. I'm also now analyzing the way the language displaces accountability from the developers to the "model" itself.

Examining the Rhetoric

I'm now focused on the rhetorical architecture. I see how the framing creates an illusion, like the curse of knowledge and the ELIZA effect. I'm exploring the premise that these systems process rather than know.

Crafting the Structure

I'm now drafting the JSON content, meticulously working to meet word count constraints while maintaining scholarly rigor. I'm focusing on incorporating specific phrasing rules, such as "this framework holds that..." and "this wording can have the effect of..." to ensure a consistent tone. I'm also ensuring I'm evaluating the language's effects, not the authors' intent. I've begun mapping mechanisms and Brown Typology instances.

Structuring the Sections

I'm now fully immersed in the structural organization, breaking down each section to align with the provided constraints. I'm focusing on hitting the word count targets for the arguments, responsibilities, Task 1-4 instances, and the conclusion. Specifically, I'm refraining from including any literal newline characters or escape sequences to ensure a strictly compliant output with the instructions. I'm now developing continuous prose to represent separate paragraphs and distinct sections within strings.

Analysis Provenance

Run ID: 2026-10-27-smmu-benchmarking-social-intelligence-of-metaphor-7sbwxi Raw JSON: 2026-10-27-smmu-benchmarking-social-intelligence-of-metaphor-7sbwxi.json Framework: Metaphor Analysis v6.8 Schema Version: 3.4.3 Generated: 2026-10-27T09:13:03.844Z

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