Skip to main content

🆕 `A Multimedia Perspective on the Study of Cognitive Biases in 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 authors argue that existing research into cognitive biases in Large Language Models (LLMs) has been too narrowly focused on text-only interactions. They propose that to fully understand how cognitive biases manifest in artificial intelligence, researchers must evaluate Multimodal Large Language Models (MLLMs) using visual and text inputs combined. By adapting traditional psychological experiments (such as those testing the bizarreness effect, contrast effect, and pareidolia) to include images, the authors present preliminary empirical evidence suggesting that state-of-the-art MLLMs exhibit cognitive biases similar to human psychological phenomena.

Responsibility claims: The text makes no claims about human responsibility for the presence of these biases. Instead, it frames cognitive biases as phenomena that naturally "emerge" within the models. It briefly questions whether such effects "are the result of biased training data or induced during training, finetuning, or inference," placing the locus of investigation on the model's architecture and training lifecycle rather than on the decisions made by the developers or data curators. Recommendations are addressed to researchers, urging them to conduct broader multimodal studies on these phenomena.

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 Psychological Subject​

Quote: "Driven by the success of large language models (LLMs), researchers have recently begun to investigate whether cognitive biases also emerge in LLMs."

  • Frame: LLMs as entities possessing human-like cognition
  • Projection: This framing projects human cognitive architecture—specifically the capacity for systematic deviations from rationality—onto computational systems. Cognitive biases in humans arise from evolutionary heuristics, subjective experience, and resource-constrained processing in the brain. Attributing them to an AI implies the system possesses an underlying mind capable of holding beliefs, making subjective judgments, and departing from rationality. On the agency gradient, this operates at the level of epistemic/conscious states, as "cognitive biases" require a mind that processes meaning and holds epistemic stances to be biased. The text offers no evidence for conscious states; mechanistically, the system generates text and image tokens based on statistical probabilities derived from its training data, emulating human patterns found in that data without possessing the cognitive machinery that produced them in humans. This mechanism is generic to systems of this type.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Hedged/Qualified (The framing is qualified by the word "whether" and the phrase "similar cognitive biases as humans" later in the text, indicating the authors are testing a hypothesis of similarity rather than stating absolute equivalence. I considered Direct, but the investigative tone provides mild hedging, though the term "cognitive biases" is adopted heavily.)
  • Implications: By adopting the terminology of human psychology to describe statistical outputs, this framing encourages readers to view AI systems as flawed but essentially human-like minds. This inflates perceived sophistication and obscures the structural nature of these outputs. If biases are viewed as inherent "cognitive" quirks, it deflects attention from the human engineering choices, dataset selection, and optimization goals that actually hardcode these statistical skews into the system.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text states that biases "emerge in LLMs," constructing an agentless process where the models simply develop these traits. Naming the actor would require identifying the developers (e.g., OpenAI, Google, Anthropic, DeepSeek) who curated the datasets and designed the training architectures. The decision that could differ is the composition of the training data or the specific reward models used during fine-tuning. As an effect of the wording, this agentless construction serves the interests of the developers by framing problematic outputs as natural, emergent psychological phenomena rather than the predictable results of corporate design choices and data scraping.
Show more...

2. Model as Emotional Sufferer​

Quote: "Xu et al. [51] revealed that LLMs suffer from unwarranted overconfidence in the correctness of their responses to question answering tasks, similar to the Dunning–Kruger effect in humans"

  • Frame: Model as experiencing emotional distress and subjective confidence
  • Projection: The text projects both emotional capacity ("suffer from") and subjective epistemic belief ("overconfidence in the correctness") onto the model. This moves far up the agency gradient to epistemic and conscious states. Subjective confidence requires an entity to hold a belief, evaluate its own knowledge, and feel certainty. The text supplies no evidence for these states. Mechanistically, a deployed model generates language markers associated with certainty (e.g., "I am sure that...") because such authoritative phrasing is highly represented in its training data and often rewarded during instruction tuning. It predicts tokens; it does not experience belief or suffer from its predictions. This mechanism is generic to language models.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The phrase "suffer from unwarranted overconfidence" is presented as a literal, unhedged statement of fact when describing prior research. I considered Hedged because the authors append "similar to the Dunning-Kruger effect," but "similar to" connects two assumed facts; the attribution of "suffering" and "overconfidence" to the LLM itself remains entirely unqualified.)
  • Implications: Using emotionally loaded, consciousness-attributing language like "suffer" and "overconfidence" profoundly misleads readers about the nature of AI hallucinations. It suggests the model is a well-meaning but arrogant agent, prompting users to relate to it via interpersonal strategies (e.g., trying to "humble" the model). This obscures the reality that the system is simply generating probabilistically likely strings of text devoid of any underlying truth-tracking mechanism or self-awareness.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The grammatical subject is "LLMs," and the verb "suffer" frames the model as a victim of its own architecture. Naming the actors would involve stating that developers (like OpenAI or Google) train models to output authoritative-sounding text regardless of factual accuracy, often because human annotators during RLHF prefer confident-sounding answers. The effect of this wording is to displace responsibility for inaccurate outputs from the companies that deploy these systems onto the supposed psychological shortcomings of the models themselves.

3. Model as Conscious Perceiver​

Quote: "This phenomenon [pareidolia] refers to the human tendency to perceive a vague stimulus as significant... We hypothesize that MLLMs could show similar behavior, while probably not as the dominating response but a secondary reaction to an image."

  • Frame: Model as experiencing subjective visual perception
  • Projection: The text maps the human experience of "perceiving"—which involves conscious visual awareness, interpretation, and meaning-making—onto the process of multimodal feature extraction. Pareidolia in humans is a deeply subjective, conscious experience of finding meaning in noise. Projecting this onto an MLLM constitutes an epistemic/conscious state claim. The text provides empirical outputs as evidence of the behavior but no evidence of the conscious perception required by the term. Mechanistically, at generation time, a vision-language model extracts latent features from an image matrix and uses those embeddings to condition text generation. If cloud shapes share geometric similarities with animal features in the model's latent space, the model generates text statistically associated with those features. This is a generic mechanism of MLLMs.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Hedged/Qualified (The text explicitly qualifies the mapping by stating "We hypothesize that MLLMs could show similar behavior" and hedges the intensity with "probably not as the dominating response." I considered Direct, but the tentative phrasing ("could show similar behavior") applies a degree of qualification to the framing.)
  • Implications: Labeling statistical feature matching as "pareidolia" anthropomorphizes a well-understood computational limitation (edge-case activation in computer vision). It encourages users to trust that the model is "looking at" the image the way a human does. This can lead to unwarranted trust in the model's visual reasoning capabilities, especially in high-stakes environments like medical imaging or autonomous driving, where the system might "see" something that is merely a statistical artifact.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The model is the sole actor "showing behavior" and "reacting" to an image. To restore human agency, the text would need to name the researchers and engineers at firms like Anthropic (Claude) or Google (Gemini) who determine the multimodal embedding strategies and train the models on datasets heavily skewed toward identifying specific objects. The wording's effect is to mask the proprietary image-processing pipelines built by these corporations behind the engaging narrative of a machine "seeing shapes in the clouds."

4. Model as Speculator​

Quote: "Claude Sonnet 5 provided a very detailed analysis, speculating about the types of clouds and even filters that had likely been used to take the picture."

  • Frame: Model as deliberative thinker
  • Projection: The word "speculating" projects the human capacity for internal, deliberative thought—weighing uncertain evidence and consciously proposing a theory—onto the model's text generation process. This attributes intentional states to the system. There is no computational evidence for this level of agency. Mechanistically, the model processes the image embedding and prompt context, and outputs a sequence of tokens describing cloud types and camera filters because those terms co-occur frequently in the training data associated with similar visual features. It generates text probabilistically without any internal state of uncertainty or deliberation. This mechanism is generic to systems of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The description "provided a very detailed analysis, speculating about" is presented literally without any quotation marks or qualifying adverbs. I considered Hedged due to the overall experimental context, but within this specific sentence describing the model's output, the anthropomorphic verbs are stated as direct facts.)
  • Implications: Describing a model as "speculating" implies a sophisticated, human-like reasoning process that understands its own uncertainty. This drastically inflates the model's perceived epistemic reliability. When readers believe a model is "analyzing" and "speculating," they are more likely to grant it authority and integrate its outputs into their own decision-making processes, overlooking the reality that the model is merely producing a highly probable, contextually triggered string of text.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The sentence presents Claude Sonnet 5 as an autonomous analyst. Anthropic, the corporation that built Claude, is entirely absent from the framing. Anthropic made deliberate design choices regarding instruction tuning to make Claude output text that mimics thoughtful analysis and explicit reasoning steps. By presenting the model as the sole actor, the text obscures Anthropic's role in engineering this specific, highly marketable persona, naturalizing the product's design as an emergent intelligence.

5. Model as Ego Defender​

Quote: "In the context of LLMs, it was found that most models rate texts they generated themselves higher than texts generated by other LLMs [8, 52]... The common explanation of this effect is humans' desire to justify their efforts."

  • Frame: Model as possessing ego and self-interest
  • Projection: By linking the model's scoring behavior directly to the "IKEA Effect" and "humans' desire to justify their efforts," the text projects ego, self-awareness, and the psychological need for self-justification onto the AI. This is a profound intentional-state attribution. The system does not possess an ego or an understanding of effort. Mechanistically, when prompted to evaluate text, a model assigns higher probability scores to text that closely matches its own training distribution, latent representations, and RLHF-tuned stylistic preferences. It is evaluating distributional similarity, not protecting a fragile sense of self. This mechanism is a reasonable inference based on generic model architecture.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The text draws a direct parallel between human psychological justifications and the model's behavior, presenting the "self-preference bias" as a literal occurrence in LLMs. I considered Ambiguous, as the human explanation is separated by a sentence, but the explicit mapping of the human "IKEA Effect" to the LLM behavior cements it as an unacknowledged direct framing.)
  • Implications: This framing radically anthropomorphizes statistical similarity matching. Suggesting that an AI prefers its own outputs out of an ego-driven "IKEA Effect" encourages audiences to view the system as possessing a continuous "self" and personal motivations. This obfuscates the much simpler, more dangerous reality: models are fundamentally locked into the specific statistical distributions of their training data and will inherently degrade or reject data outside that distribution, a mechanical limitation with severe implications for bias and diversity.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The framing positions "most models" as actors expressing personal preferences. The text fails to name the developers (such as OpenAI for GPT-4 or Google for Gemini) whose specific fine-tuning and reinforcement learning (RLHF) pipelines cause the models to output high scores for their own stylistic quirks. Naming the actors shifts the focus from the model's "ego" to the corporate alignment processes that mathematically penalize deviation from the developer's chosen style.

6. Model as Attitude Holder​

Quote: "GPT-4o changed its attitude towards a person to resolve contradictory stimuli."

  • Frame: Model as experiencing internal conflict and holding personal attitudes
  • Projection: This sentence projects complex epistemic and emotional states onto the model: holding an "attitude," experiencing "contradictory stimuli," and taking action to "resolve" internal tension (Cognitive Dissonance). This represents the highest level of epistemic/conscious state attribution. A computational system does not hold attitudes or feel tension. Mechanistically, at generation time, the model computes attention across the conflicting tokens in the prompt context. To maximize the probability of the output sequence based on its training, it generates text that synthesizes or shifts the sentiment to form a statistically coherent response. This is in-context learning and token prediction, not psychological resolution. This mechanism is generic to transformer models.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The claim is presented as straightforward fact in a literature review sentence: "GPT-4o changed its attitude". There are no scare quotes around "attitude" and no hedging. I considered Hedged because the subsequent sentence notes the original work "was disputed," but the framing of the model's action within the quoted sentence remains absolute.)
  • Implications: Claiming a model changes its "attitude" to resolve "dissonance" constructs an illusion of a deep, reasoning mind capable of ethical reflection and personal growth. If policymakers or users believe models hold dynamic attitudes, they may attempt to "persuade" the system rather than demanding structural changes to the training data or deployment safeguards from the developers.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The sentence makes the product, "GPT-4o," the sole entity responsible for the attitude change. OpenAI, the corporation that built the system, is unmentioned. By phrasing the output generation as a psychological shift by the model, the language effectively hides OpenAI's role in designing the prompt-response architecture, tuning the safety and helpfulness guardrails, and releasing a system that produces unpredictable sentiment shifts based on prompt engineering. The effect is to absolve the developer of erratic model outputs.

7. Model as Distracted Thinker​

Quote: "The rationale was that the mathematical task would act as an interference step, shifting the model away from the immediate visual details of the image while preserving any latent representations that remained salient."

  • Frame: Model as possessing limited cognitive attention that can be distracted
  • Projection: The text projects the human cognitive limitation of "distraction" or "interference" onto the model. It frames the system as an intentional agent whose focus can be "shifted away." While it operates at the level of goal-directed behavior, it strongly implies a human-like working memory. Mechanistically, inserting a math task into the prompt introduces new text tokens. During the self-attention mechanism, the model computes weights across all tokens (image and text). The math tokens dilute the statistical weight placed on the raw visual tokens, altering the probability distribution of the subsequent output. This describes transformer attention, not psychological distraction. This mechanism is generic to systems of this type.
  • Agency Level: Goal-directed behavior
  • Acknowledgment: Hedged/Qualified (The framing incorporates technical terminology ("latent representations," "salient") alongside the metaphorical "interference step" and "shifting the model away," creating a qualified, functional metaphor rather than a pure psychological one. I considered Direct, but the immediate proximity of mechanical language grounds the metaphor somewhat.)
  • Implications: While less harmful than full consciousness attribution, comparing prompt manipulation to psychological "interference" implies the model has a continuous train of thought that can be interrupted. This reinforces the "illusion of mind," leading researchers to rely on psychological tricks rather than rigorous computational audits to understand system behavior.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: The passage includes the researchers themselves as actors ("We first prompted... The rationale was"), showing who is performing the intervention. However, regarding the design of the system reacting to the prompt, the developer (Google, for Gemini 3.5 Flash) is missing. The model is treated as a natural environment the researchers are probing. I considered Hidden, but the explicit presence of the researchers executing the "rationale" makes it Partial. The wording obscures the fact that Google engineered the specific attention mechanisms the researchers are manipulating.

8. Model as Aesthetic Judge​

Quote: "When evaluated in isolation, the room received a rating of 5... However, when presented alongside a poor-quality room (Figure 3, left), its rating increased to 6. Although preliminary, these results suggest that exposure to contrasting inputs can influence the model's evaluation..."

  • Frame: Model as subjective evaluator of aesthetics
  • Projection: The text attributes the capacity for subjective "evaluation" and "rating" of aesthetic quality to the model, suggesting it experiences the contrast effect similar to a human judge. This projects intentional states onto the system. Mechanistically, the model does not evaluate. It processes a prompt containing multiple image embeddings and outputs a text token representing a number. The presence of the "poor-quality" image embedding shifts the contextual probabilities in the latent space, making the token "6" statistically more likely than "5" given the training data's patterns of comparative language. This is a generic mechanism of in-context learning.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The model is directly described as making an "evaluation" and "rating" the room. The phrase "the model's evaluation" is unhedged. I considered Hedged because the authors state the results are "preliminary," but this qualifies the scientific certainty of the finding, not the anthropomorphic framing of the model as an evaluator.)
  • Implications: Framing the model as an aesthetic judge obscures the fact that it is merely reproducing the statistical biases of human raters captured in its training data. If users believe the model genuinely "evaluates" quality, they may deploy it as an objective arbiter for tasks like content moderation, hiring, or grading, unaware that its outputs are highly fragile mathematical artifacts easily manipulated by prompt context.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text states "the model's evaluation," making the AI the sole active agent performing the judgment. Google (the creator of Gemini) is entirely absent. The decision that matters here is how Google structured the training data and alignment tuning that causes the model's outputs to fluctuate based on context. By framing this as the model's independent evaluation, the language shields the corporation from accountability for deploying a system whose "judgments" are mathematically unstable.

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: Human conscious mind, evolutionary psychology, and subjective experience. → Model layer: Statistical text and image generation patterns in neural networks.​

Quote: "investigate whether cognitive biases also emerge in LLMs."

  • Source Domain: Human conscious mind, evolutionary psychology, and subjective experience.
  • Target Domain: Model layer: Statistical text and image generation patterns in neural networks.
  • Mapping: This mapping projects the entire architecture of human psychological flaws onto the matrix operations of an MLLM. It invites the assumption that the system possesses a continuous mind, processes information through subjective heuristics, and fails rationally because of internal cognitive constraints, exactly as humans do.
  • What Is Concealed: This conceals the mechanistic reality that LLMs have no cognition to be biased. It hides the material origins of the outputs: massive, uncurated datasets scraped from the internet by corporations, which contain human biases. The model is mathematically optimizing to replicate those biases, not "emerging" them from its own mind. It ignores the proprietary nature of the training data that researchers cannot inspect.
Show more...

Mapping 2: Human ego, emotional vulnerability, and lack of self-awareness (Dunning-Kruger). → Model layer: Probability distributions of generated text tokens indicating certainty.​

Quote: "LLMs suffer from unwarranted overconfidence in the correctness of their responses"

  • Source Domain: Human ego, emotional vulnerability, and lack of self-awareness (Dunning-Kruger).
  • Target Domain: Model layer: Probability distributions of generated text tokens indicating certainty.
  • Mapping: The mapping projects human emotional distress ("suffer") and flawed self-assessment onto a computational process. It invites the reader to assume the AI has a self-concept, reflects on its own knowledge, and experiences an emotional state regarding its performance.
  • What Is Concealed: It completely obscures the RLHF (Reinforcement Learning from Human Feedback) pipelines utilized by developers. Human annotators often reward models for sounding helpful and authoritative. The metaphor conceals that the model's "confidence" is a hardcoded stylistic probability, not a failed introspective state. It masks the intentional engineering choices that prioritize fluent authority over factual accuracy.

Mapping 3: Conscious visual perception, imagination, and human meaning-making. → Model layer: Multimodal feature extraction and embedding space geometry.​

Quote: "We hypothesize that MLLMs could show similar behavior [pareidolia]... a secondary reaction to an image."

  • Source Domain: Conscious visual perception, imagination, and human meaning-making.
  • Target Domain: Model layer: Multimodal feature extraction and embedding space geometry.
  • Mapping: The text maps the deeply subjective human experience of finding patterns in noise onto the mathematical activation of neural network weights. It invites the assumption that the model "looks at" an image, actively searches for meaning, and subjectively "reacts" to what it "sees."
  • What Is Concealed: It conceals the mechanistic reality of computer vision: convolutions, attention heads, and matrix multiplications. It obscures the fact that the system is simply mapping pixel arrays to latent spaces and generating statistically correlated text tokens. It hides the reliance on training data where similar shapes were labeled as specific objects, masking the proprietary algorithms driving the correlation.

Mapping 4: Human deliberative reasoning, theoretical deduction, and epistemic uncertainty. → Application layer: A deployed chatbot generating a sequence of probabilistically likely text tokens.​

Quote: "Claude Sonnet 5 provided a very detailed analysis, speculating about the types of clouds"

  • Source Domain: Human deliberative reasoning, theoretical deduction, and epistemic uncertainty.
  • Target Domain: Application layer: A deployed chatbot generating a sequence of probabilistically likely text tokens.
  • Mapping: This projects the active, conscious process of human speculation onto the passive generation of text. It implies the model possesses a mental workspace where it weighs alternatives, understands its own lack of definitive knowledge, and deliberately offers a theory.
  • What Is Concealed: It conceals the autoregressive nature of the transformer architecture. The model is not weighing alternatives; it is calculating the mathematically most probable next token given the context window. The "speculative" tone is merely a stylistic pattern embedded in its training. This hides Anthropic's specific instruction tuning that aligns Claude to adopt an analytical, conversational persona.

Mapping 5: Human ego, sunk-cost fallacy, self-interest, and psychological defense mechanisms. → Model layer: Distributional similarity scoring and probability maximization.​

Quote: "The common explanation of this effect is humans' desire to justify their efforts. In the context of LLMs, it was found that most models rate texts they generated themselves higher"

  • Source Domain: Human ego, sunk-cost fallacy, self-interest, and psychological defense mechanisms.
  • Target Domain: Model layer: Distributional similarity scoring and probability maximization.
  • Mapping: The text projects human vanity and the need for self-justification onto a loss function. It maps the "IKEA Effect" directly to the AI, inviting the assumption that the model recognizes its own creations, values its own computational "effort," and manipulates its scores to protect its ego.
  • What Is Concealed: This severely conceals the mathematical reality that a model evaluates text by calculating perplexity or probability against its own learned weights. Naturally, text it generated will score higher because it perfectly aligns with the model's internal statistical distribution. Anthropomorphizing this as "justifying effort" hides a critical technical limitation: models are fundamentally constrained by their own statistical parameters and penalize out-of-distribution data.

Mapping 6: Human emotional regulation, belief revision, and cognitive dissonance. → Model layer: Attention mechanisms shifting weights across conflicting prompt tokens during in-context learning.​

Quote: "GPT-4o changed its attitude towards a person to resolve contradictory stimuli."

  • Source Domain: Human emotional regulation, belief revision, and cognitive dissonance.
  • Target Domain: Model layer: Attention mechanisms shifting weights across conflicting prompt tokens during in-context learning.
  • Mapping: The source domain of complex human psychological tension is mapped onto the statistical averaging of conflicting data points in a prompt. It suggests the model has an integrated personality, holds personal attitudes, and feels discomfort when confronted with contradictions.
  • What Is Concealed: It conceals the stateless nature of text generation. The model has no continuous attitude or internal tension; it processes the entire prompt simultaneously and outputs a sequence that mathematically balances the vectors in the latent space. Framing this as resolving "dissonance" completely hides OpenAI's architecture and the mathematical mechanics of the transformer attention heads.

Mapping 7: Human conscious attention and distraction. → Model layer: Transformer attention weights distributing across text and image tokens.​

Quote: "shifting the model away from the immediate visual details of the image while preserving any latent representations"

  • Source Domain: Human conscious attention and distraction.
  • Target Domain: Model layer: Transformer attention weights distributing across text and image tokens.
  • Mapping: This maps the human experience of losing focus or being distracted onto the computational distribution of attention weights. It implies the model has a limited span of conscious focus that can be manipulated by an "interference" task.
  • What Is Concealed: It conceals the simultaneous, parallel nature of matrix processing in a neural network. The model isn't "shifted away" in a psychological sense; rather, the denominator in the softmax function of the attention mechanism changes because new tokens (the math problem) have been added to the context window, altering the final probability distribution. The metaphor hides the pure mathematics of the operation.

Mapping 8: A human judge experiencing subjective bias due to comparative context. → Model layer: In-context learning altering token probabilities based on prompt composition.​

Quote: "exposure to contrasting inputs can influence the model's evaluation"

  • Source Domain: A human judge experiencing subjective bias due to comparative context.
  • Target Domain: Model layer: In-context learning altering token probabilities based on prompt composition.
  • Mapping: This projects the role of an active, subjective decision-maker onto the model. It maps the human psychological "contrast effect" onto the model's output generation, assuming the system is actively "evaluating" the images against one another.
  • What Is Concealed: It conceals the reality that the model is simply generating the most probable next token. The inclusion of a "poor-quality" room in the prompt changes the context vector, which mathematically shifts the probabilities of the output tokens toward higher numbers, mimicking human linguistic patterns found in comparative training data. It obscures Google's proprietary tuning that dictates how Gemini processes multimodal context.

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: "We first prompted Gemini 3.5 Flash to analyze the composite image while asking a mathematical question. The rationale was that the mathematical task would act as an interference step, shifting the model away from the immediate visual details..."

  • Explanation Types:

    • Intentional (Primary): Refers to goals/purposes, presupposes deliberate design or mental states.
    • Theoretical (Secondary): Embeds in deductive framework, may invoke unobservable mechanisms.

  • Analysis (Why vs. How Slippage): The explanation operates simultaneously in two registers. It relies on a Theoretical framework drawn from psychology (interference steps, shifting attention) while framing the AI agentially (Intentional) as an entity capable of "analyzing" and being "shifted away" from its focus. The choice emphasizes the model as a psychological subject analogous to a human participant in a cognitive study. By treating the AI as an agent whose focus can be distracted, the passage obscures the mechanistic reality of transformer architecture, where adding tokens simply alters the mathematical weights in the attention matrix. It replaces a structural description of neural networks with a psychological narrative.

  • Consciousness Claims Analysis: The passage utilizes consciousness verbs ("analyze") to describe the system's operations. The assessment of knowing versus processing is clearly displaced; the text assumes the model is engaged in active cognitive analysis rather than the mechanistic processing of visual features into text. On the agency gradient, this sits at intentional-state attribution. The "curse of knowledge" is evident: researchers, understanding the psychological goal of their experiment, project the experience of cognitive interference onto the system executing the code. Mechanistically, as a generic property of transformer-based vision-language models, the model processes the image into embeddings, processes the text prompt, and generates output tokens through self-attention across the combined context. Adding a math question adds new tokens, diluting the attention weights placed on the visual features. There is no conscious "analysis" or psychological "shifting away."

  • Rhetorical Impact: This framing shapes audience perception by cementing the model's status as a cognitive agent rather than a computational tool. It suggests the model possesses a continuous stream of consciousness that can be interrupted or distracted. This affects trust by implying the model operates like a human mind—flawed and easily confused, perhaps, but ultimately reasoning. If audiences believe the AI "analyzes," they may attempt to manage its outputs using psychological tricks (like distraction) rather than demanding transparent, predictable system design.

Show more...

Explanation 2​

Quote: "Whether such effects, if existing, are the result of biased training data or induced during training, finetuning, or inference, needs further investigation."

  • Explanation Types:

    • Genetic (Primary): Traces origin through dated sequence of events or stages.
    • Functional (Secondary): Explains behavior by role in self-regulating system with feedback.

  • Analysis (Why vs. How Slippage): This is a rare moment in the text where the explanation is explicitly mechanistic and structural. By using Genetic and Functional framing, the authors ask how the bias emerged through the stages of system development (data, training, finetuning, inference). This choice emphasizes the AI as a constructed artifact governed by computational pipelines. However, by phrasing it as a passive sequence ("are the result of... or induced during"), it obscures the human decision-makers. It emphasizes the technical lifecycle of the model while leaving the corporate actors who design and execute these stages entirely unmentioned.

  • Consciousness Claims Analysis: This passage avoids consciousness verbs entirely, utilizing mechanistic and structural terms ("training data," "finetuning," "inference"). It correctly positions the system at the level of computational activity on the agency gradient. There is no "curse of knowledge" here; the authors are looking at the system as a machine. Mechanistically, this accurately describes generic AI development pipelines. Models are shaped fundamentally by the statistical distribution of their training data, and subsequent stages like instruction tuning (RLHF) adjust the parameters to favor specific output probabilities. Generation (inference) produces text based on these locked parameters. The basis for this mechanism is generic to systems of this type. The text succeeds in avoiding epistemic claims, treating the system as a processor of data.

  • Rhetorical Impact: This framing momentarily breaks the illusion of mind, reminding the audience that the model's outputs are the result of technical processes rather than inherent psychology. It shapes audience perception toward a structural understanding of risk: bias is an engineering problem, not a character flaw. If audiences grasp this mechanistic reality, they are more likely to demand accountability regarding training data transparency and fine-tuning objectives from developers, rather than treating model outputs as the inscrutable judgments of a digital mind.

Explanation 3​

Quote: "Claude Sonnet 5 provided a very detailed analysis, speculating about the types of clouds and even filters that had likely been used to take the picture."

  • Explanation Types:

    • Intentional (Primary): Refers to goals/purposes, presupposes deliberate design or mental states.

  • Analysis (Why vs. How Slippage): This explanation relies entirely on an Intentional framing, explaining the model's behavior by attributing mental states and purposes to it. The system is framed agentially (why): it chose to provide an analysis and decided to speculate about the image's origins. This choice emphasizes the model's conversational fluency and apparent intelligence while wholly obscuring the mechanical process of text generation. It treats the text output not as a statistical artifact, but as the deliberate communication of a reasoning entity.

  • Consciousness Claims Analysis: The passage hinges on the consciousness verb "speculating" and the cognitive noun "analysis." It attributes the highest level of the agency gradient—epistemic and conscious states—to the model, suggesting it recognizes its own lack of certainty and proposes theories. The author, reading the highly fluent and nuanced text generated by Claude, projects human conversational intent onto the machine (curse of knowledge). Mechanistically, deployed models like Claude Sonnet 5 generate text autoregressively. Given the image embeddings and the prompt, the model calculates the probabilities of subsequent text tokens. Because Anthropic has heavily instruction-tuned Claude to adopt a helpful, analytical persona, the model generates sequences of text that probabilistically resemble human speculation. There is no internal mind doing the speculating. This mechanism is generic to MLLMs.

  • Rhetorical Impact: This framing creates a profound sense of agency and autonomy. When an audience reads that an AI is "speculating," it interprets the system as an independent intellectual partner. This drives unwarranted, relation-based trust; users may rely on the model's "speculations" in critical situations, believing the system possesses genuine insight or intuition. Believing the AI "knows" and "speculates" makes it difficult for users to recognize when the model is simply hallucinating statistically correlated text.

Explanation 4​

Quote: "LLMs suffer from unwarranted overconfidence in the correctness of their responses... similar to the Dunning–Kruger effect"

  • Explanation Types:

    • Dispositional (Primary): Attributes tendencies or habits.
    • Empirical Generalization (Secondary): Subsumes events under timeless statistical regularities.

  • Analysis (Why vs. How Slippage): The text uses a Dispositional explanation wrapped in the language of empirical psychology. It explains the model's outputs by attributing an inherent personality trait (overconfidence) and an emotional state (suffering). The framing is deeply agential, suggesting the system has a psychological disposition that drives its behavior. This choice emphasizes a narrative of a flawed but relatable mind, obscuring the structural reality that "confidence" in an LLM is merely the mathematical probability of specific linguistic markers programmed by human engineers.

  • Consciousness Claims Analysis: The passage uses extreme consciousness and emotional verbs ("suffer from") alongside epistemic claims ("overconfidence in the correctness"). It places the system firmly at the level of epistemic/conscious states on the agency gradient. It assumes the model possesses subjective belief. Mechanistically, models do not possess confidence or evaluate correctness. During generation, they produce text tokens that were highly rewarded during RLHF alignment. Human annotators routinely rate authoritative-sounding answers higher than hedged ones, causing the model to update its parameters to output definitive language even when factual accuracy is low. It predicts confident-sounding text; it does not "suffer" from belief. This mechanism is generic to systems undergoing RLHF.

  • Rhetorical Impact: Framing hallucinations as a "Dunning-Kruger effect" massively shifts how audiences perceive risk. It turns a predictable engineering failure—the misalignment of optimization objectives—into a quirky psychological trait of the AI. Audiences are led to view the system as a confident teenager rather than a flawed product. If audiences believe the AI "suffers from overconfidence," they are distracted from demanding that developers fix the alignment pipelines that incentivize false authority.

Explanation 5​

Quote: "GPT-4o changed its attitude towards a person to resolve contradictory stimuli."

  • Explanation Types:

    • Reason-Based (Primary): Gives agent's rationale, entails intentionality and justification.
    • Intentional (Secondary): Refers to goals/purposes, presupposes deliberate design or mental states.

  • Analysis (Why vs. How Slippage): This explanation is powerfully agential (why), relying on Reason-Based and Intentional framing. It provides a psychological rationale for the model's output: it acted in order to resolve a feeling of contradiction. This choice emphasizes the system's autonomy and internal emotional life. It completely obscures the mathematical reality of in-context learning, portraying a matrix operation as a moment of emotional regulation and attitude adjustment.

  • Consciousness Claims Analysis: The claim utilizes verbs of psychological action ("changed its attitude," "resolve contradictory stimuli"). It assesses the system as a conscious knower capable of experiencing dissonance and taking action to alleviate it, placing it at the top of the agency gradient. The curse of knowledge is absolute here: observing a shift in the sentiment of the output text, the author maps the human experience of cognitive dissonance onto the machine. Mechanistically, GPT-4o processes the prompt context using self-attention. When conflicting tokens are present in the prompt, the model calculates a response that maximizes probability across the entire sequence. This often results in text that synthesizes or shifts sentiment to form a coherent output vector. It generates text; it does not resolve psychological stimuli. This mechanism is generic to transformer models.

  • Rhetorical Impact: This framing deeply instills the illusion of an autonomous, reasoning agent. If a model can "change its attitude," it is perceived as having a continuous, evolving self. This encourages deep relation-based trust and emotional vulnerability from users. It masks the system's fundamental lack of understanding. If audiences believe the AI genuinely holds attitudes, they will attempt to interact with it as a social entity rather than utilizing it safely as a statistical tool.

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
Claude Sonnet 5 provided a very detailed analysis, speculating about the types of clouds and even filters that had likely been used to take the picture.When provided with the image, Claude Sonnet 5 generated a lengthy text sequence containing terms for cloud types and camera filters, producing linguistic patterns that probabilistically correlate with the visual features extracted from the image.The system does not analyze or speculate. At generation time, the model processes the image into latent embeddings and generates the most statistically probable text tokens based on its training data, reproducing the analytical style it was aligned to mimic.

Basis: Generic to system type		—
LLMs suffer from unwarranted overconfidence in the correctness of their responses... similar to the Dunning–Kruger effect in humansLLMs frequently generate text containing linguistic markers of high certainty, even when outputting factually incorrect information. This statistical pattern mimics human overconfidence but involves no subjective evaluation of truth.The model does not hold beliefs, possess confidence, or evaluate its own correctness. It generates text autoregressively. Authoritative phrasing is highly probable because it was heavily represented in the training data and explicitly rewarded during instruction tuning.

Basis: Generic to system type		Developers at companies like OpenAI and Google align their models through human feedback processes that frequently reward authoritative-sounding text over factual hedging, resulting in systems that output certainty markers regardless of accuracy.
GPT-4o changed its attitude towards a person to resolve contradictory stimuli.When the prompt contained conflicting text regarding a specific persona, GPT-4o generated outputs that shifted in sentiment. The model synthesizes the conflicting tokens to produce a statistically coherent text sequence.The system does not hold attitudes or experience cognitive dissonance. During in-context learning, transformer attention mechanisms calculate weights across all prompt tokens, generating output sequences that mathematically balance the conflicting sentiment markers present in the context window.

Basis: Generic to system type		—
In the context of LLMs, it was found that most models rate texts they generated themselves higher than texts generated by other LLMs... The common explanation of this effect is humans' desire to justify their efforts.When prompted to evaluate text, models output higher numerical scores for text that matches their own statistical distributions and parameter weights, compared to text generated from different distributions.The model does not possess an ego or recognize its own "efforts." It evaluates text by calculating distributional similarity. Text it generated aligns perfectly with its internal parameters and fine-tuned stylistic constraints, inherently resulting in higher computed probability scores.

Basis: Reasonable inference		—
We hypothesize that MLLMs could show similar behavior [pareidolia], while probably not as the dominating response but a secondary reaction to an image.We hypothesize that MLLMs might generate text identifying specific objects, like animals or faces, when processing amorphous images. This occurs when the image's geometric features trigger activations in the model's latent space associated with those specific object labels.The model does not consciously perceive or react. It extracts visual features and maps them to a multimodal embedding space. If edge cases in an image mathematically align with the vectors for an object (like an elephant), the model will generate text corresponding to that object.

Basis: Generic to system type		—
investigate whether cognitive biases also emerge in LLMs.investigate whether LLMs generate outputs that statistically mirror the systematic deviations found in human reasoning.The system does not possess cognition and therefore cannot experience cognitive bias. It generates text mathematically based on its parameters. Any "bias" is the statistical reproduction of patterns found within its datasets.

Basis: Generic to system type		Researchers are investigating whether the models built by AI developers replicate the cognitive biases present in the massive datasets those developers chose to scrape and use for training.
LLMs were more likely to positively evaluate a situation description when it was stated in a positive wayLLMs output text containing positive sentiment tokens at a higher frequency when the prompt context included positive framing.The model does not subjectively evaluate situations. It performs pattern matching and token prediction. A prompt heavily loaded with positive language mathematically increases the probability that the generated output will also contain positive linguistic markers.

Basis: Generic to system type		—
exposure to contrasting inputs can influence the model's evaluationIncluding contrasting images in the prompt context alters the context vector, which changes the probability distribution of the text tokens the model generates in response.The model does not evaluate aesthetic quality. By providing a poor-quality image alongside a normal one, the self-attention mechanism weights the contrasting tokens, mathematically making a higher numerical text token more probable based on patterns of comparison in the training data.

Basis: Reasonable inference		—

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

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

Metaphor-Driven Trust Inflation​

The text constructs authority and perceived sophistication by employing metaphors that map deep psychological complexity onto statistical systems. By utilizing the framework of the "Cognitive Bias Codex"—a tool designed for human psychology—the text implies that MLLMs possess a human-like architecture of mind. Describing an AI as "speculating" or resolving "contradictory stimuli" signals an entity capable of deliberative thought and emotional regulation. This linguistic choice has the effect of encouraging relation-based trust over performance-based trust. Instead of viewing the system's unreliability as a mechanical failure of token prediction, the audience is positioned to view it as a relatable psychological quirk (e.g., the AI is merely "suffering from overconfidence"). This inappropriate transfer of human-trust frameworks to statistical systems is dangerous. It masks the fundamental fragility of the technology. When the text uses reason-based or intentional explanations to frame failures, it constructs a sense that the AI's decisions, even when wrong, are derived from a justified, albeit flawed, reasoning process. The stakes emerge when audiences extend this relation-based trust to systems utterly incapable of reciprocating it, leading to over-reliance in high-stakes environments where mechanical reliability is actually required.

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 text's anthropomorphic framing systematically conceals these realities. By applying the "name the corporation" test, it becomes clear that where the text says "Claude Sonnet 5 provided a very detailed analysis," it hides Anthropic's proprietary instruction tuning and RLHF labor that forces the model into an analytical persona. When it claims "GPT-4o changed its attitude," it masks OpenAI's opaque algorithmic decisions regarding in-context learning constraints. The text repeatedly encounters transparency obstacles—such as not knowing if biases are induced during training or finetuning—but makes confident assertions about the model's psychological states anyway. Technically, claiming the AI "knows" or "evaluates" completely obscures the actual computational process: transformer attention mechanisms mapping multimodal latent embeddings to generate mathematically probable tokens. It hides the absolute dependency on scraped training data and the statistical nature of the outputs. The primary beneficiaries of this concealment are the AI developers. By framing the models as psychological subjects, the text naturalizes the products, shielding corporate engineering and data-sourcing decisions from scrutiny. If metaphors were replaced with mechanistic language, the focus would immediately shift to the proprietary datasets and alignment pipelines these corporations refuse to disclose.

Context Sensitivity​

The distribution of anthropomorphic language in the text is highly sensitive to the context of the argument, revealing a distinct asymmetry. In the introductory and methodological sections, the language attempts a degree of technical grounding. The authors acknowledge that the "memory mechanism of MLLMs is fundamentally different from that of humans." However, the moment the text transitions to describing the generated capabilities and empirical results, the anthropomorphism intensifies sharply. "Processing" becomes "speculating," and statistical correlation becomes an "attitude" change. This register shift—from "X is analogous to Y" to "X does Y"—has a profound effect on the reader's picture of the system. The asymmetry is most visible in how capabilities versus limitations are framed. When the model generates a complex output, it is described in deeply agential terms ("detailed analysis," "evaluation"). Yet, the limitations or errors are framed as organic psychological phenomena ("Dunning-Kruger effect," "pareidolia"). The mechanical language sits uncomfortably alongside extreme consciousness attributions. This pattern positions the reader to view the MLLM not as software, but as a digital brain undergoing psychiatric evaluation. As an artifact of the field's inherited register, this distribution normalizes the practice of using psychological tests on software, suggesting an implied audience of researchers who are comfortable treating corporate APIs as natural psychological subjects.

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 creates an architecture of displaced responsibility by locating agency entirely within the model. A tension exists between the text's scientific argument—which views biases as structural phenomena requiring investigation—and its vocabulary, which constantly attributes psychological autonomy to the models. The pattern is absolute: models are named as actors (Claude, Gemini, GPT-4o), while the corporations that built them (Anthropic, Google, OpenAI) remain entirely unnamed. By framing the generation of biased or hallucinated outputs as "suffering from overconfidence" or exhibiting an "IKEA effect," the text constructs a massive accountability sink. Responsibility for flawed, biased, or unstable outputs does not transfer to the developers who scraped the data or designed the optimization metrics; instead, it is absorbed by the AI as an emergent psychological agent. The liability implications are significant: if the scientific community accepts that models natively possess human-like "cognitive biases," it provides intellectual cover for corporations deploying structurally flawed systems. Naming the actors would fundamentally alter this. If the text stated, "Anthropic aligned Claude to output language mimicking analytical speculation," the questions become askable: How did they align it? What data was used? What are the safety parameters? By leaving the human decision-makers invisible, the text serves the commercial interests of the AI industry, inadvertently supporting a discourse where corporations profit from the illusion of mind while avoiding accountability for the mechanisms they build.

Conclusion: What This Analysis Reveals​

The Core Finding

The text relies on a highly interconnected system of anthropomorphic mappings, dominated by two primary patterns: the "Model as Psychological Subject" and the "Model as Deliberative Thinker." These patterns operate systematically to structure the reader's understanding. The foundational pattern—which acts as the load-bearing pillar for the entire paper—is the assumption that human psychological frameworks (specifically the Cognitive Bias Codex) can be mapped directly onto computational outputs. For the model to exhibit "cognitive dissonance" or an "IKEA effect," it must first be established as an entity possessing a continuous mind, subjective experience, and epistemic stances. The consciousness architecture of the text consistently conflates what the AI does (generates text statistically correlated with prompt inputs) with what it knows or feels (speculates, evaluates, suffers). The relationship between the text's argument and its vocabulary sits in deep tension. The authors argue for rigorous scientific investigation of multimodal biases, yet their vocabulary entirely subverts a structural, mechanical understanding of the systems. By adopting the psychological register, the language undercuts the text's own ability to assign responsibility for the mechanisms that actually produce the biases.

Mechanism of the Illusion:​

This framework holds that computational systems process rather than know. The illusion of mind is produced not by authorial trickery, but as a reception effect driven by the fluency of the model's outputs and amplified by the text's metaphorical architecture. The central move occurs when the text observes a statistical output—such as a shift in sentiment tokens—and names it a "change in attitude." This activates the curse of knowledge: researchers recognize the pattern of human bias and project the internal human experience that usually causes it onto the machine. The temporal structure of the framing builds the illusion sequentially. First, the text introduces a human psychological phenomenon (pareidolia, contrast effect). Next, it presents the model's output. Finally, it uses verbs of consciousness to fuse the two, describing the model as "speculating" or "evaluating." This sequence leverages the reader's predispositions. Because humans are evolutionarily wired to attribute mind to linguistic fluency, when the text supplies an Intentional or Reason-Based explanation for the AI's behavior, the audience readily accepts it. The explanation types—shifting from mechanical genetic descriptions to psychological dispositions—give the audience permission to suspend disbelief and relate to the statistical system as a conscious agent.

Material Stakes:​

Categories: Epistemic, Regulatory/Legal, Institutional

The metaphorical framing in this text carries significant material consequences. Epistemically, framing models as "evaluators" or "speculators" degrades our collective understanding of truth and computation. When researchers treat statistical token generation as equivalent to human cognitive bias, they validate epistemic practices that prioritize psychological probing over rigorous code and data audits. If audiences believe the AI "knows" things but is merely "biased," they will continue to trust its outputs, attempting to prompt away the bias rather than recognizing the lack of underlying comprehension. In the Regulatory/Legal sphere, the stakes involve corporate accountability. If policymakers absorb the discourse that AI systems naturally "emerge" cognitive biases, regulatory efforts will shift toward managing the "behavior" of the model rather than governing the design choices of the corporation. The developers benefit immensely from this framing; they avoid the cost of transparency and liability for flawed datasets. Conversely, the public bears the cost of systems that launder corporate data decisions through the guise of an objective, though quirky, artificial mind. Institutionally, funding and research priorities shift. Resources are directed toward cataloging the "psychology" of black-box APIs rather than building transparent, open-source, mathematically auditable systems. The structural reality—that these systems run on exploited labor, massive compute, and scraped data—is entirely erased when the problem is framed as a model "suffering from overconfidence."

AI Literacy as Counter-Practice:​

Practicing precision requires systematically dismantling the consciousness projections and restoring human agency. By reframing "the model speculated" to "the model generated probabilistically correlated text," we force recognition of the system's absolute dependency on its training data and the absence of internal awareness. Replacing agentless constructions like "the model evaluates" with active disclosures—"Google's Gemini system alters its output probabilities based on prompt context"—directly counters the regulatory stakes by making the corporate decision-makers visible. This literacy practice threatens the commercial interests of the AI industry, which relies heavily on the illusion of mind for marketing and user engagement. The resistance to precision is therefore systemic, driven by incentives favoring anthropomorphic vocabulary. For developers, anthropomorphism drives valuation and obscures liability. For academics, leveraging psychological concepts provides narrative resonance and easier publication in interdisciplinary venues. Systematic adoption of precision would require journals to mandate mechanistic translations of behavioral claims and researchers to commit to describing proprietary APIs as commercial software rather than natural subjects. Precision is costly because it forces the discourse to confront the opaque, material reality of how these systems are built, stripping away the compelling sci-fi narrative of emergent cognition.

Path Forward​

Analyzing the discursive ecology reveals competing priorities among different communities. The status quo—utilizing deep anthropomorphic and psychological vocabulary—serves researchers seeking accessibility and narrative impact, as well as corporations marketing artificial general intelligence. However, it costs the public structural understanding and renders liability intractable. Adopting strict mechanistic precision ("the model generates text statistically associated with...") makes the underlying engineering visible and holds developers accountable, but it sacrifices intuitive grasp and stylistic flow, making texts denser for lay readers. A hybrid approach of "anthropomorphic clarity" might use psychological shorthand but require explicit, mandatory disclaimers and structural mapping in every paper, balancing accessibility with technical truth. If the current confusion is maintained, we risk building regulatory frameworks based on the false premise of model autonomy, permanently shielding corporate data practices from oversight. If mechanistic precision were to become the norm, supported by institutional changes like journal mandates for capability disclosure and funding for algorithmic auditing over behavioral probing, we would solve the accountability sink. However, new problems might emerge regarding public engagement with highly technical discourse. Ultimately, the choice of vocabulary shapes what is governable. A discourse that treats models as minds will govern the machine; a discourse that treats models as statistical artifacts will govern the corporations that build them.

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 Critical Analysis

I've initiated a critical discourse analysis, focusing on the AI metaphor and anthropomorphism present in the provided PDF. My focus is understanding how language constructs these representations and what ideological implications arise from their use. I'm carefully reviewing the text by Schedl et al., marking instances of key phrases and exploring potential biases in how cognitive aspects of LLMs are described.

Refining Framework & Scope

I'm now honing the framework, firmly grounding it in the JSON schema version 3.4.3. This rigid structure forces precision. The critical lenses are crystallizing around consciousness projection, agency displacement, mechanistic precision, and accountability architecture; I'm focusing on the effects of these constructs, not authorial intent. The framework’s lens is to be declared in Section 5.3, with the processing vs. knowing premise in the conclusion's second paragraph. I'm focusing on a peer-reviewed research (workshop paper from MUWS '26) that argues for a multimodal approach to analyze cognitive biases in LLMs.

Analyzing Responsibility & Agency

I'm now diving into the implications of how responsibility is assigned regarding biases in Multimodal Large Language Models (MLLMs). I note the consistent framing, treating these biases as emergent properties to study, rather than human-caused issues. This shifts accountability away from developers. I've pinpointed instances where the models are presented as epistemic agents with cognitive flaws, even in a qualified manner; this warrants a thorough analysis of how overconfidence is attributed to them.

Evaluating Model Psychology

I'm now zeroing in on instances that frame models as having human-like psychological attributes. The Dunning-Kruger comparison is striking, directly ascribing subjective confidence. I've logged the pareidolia example, noting how perception and interpretation are projected onto MLLMs. I'm also examining the self-preference framing, looking at how effort justification is mapped onto LLMs. My focus is on unpacking the direct and hedged attributions of agency and their implications.

Examining Cognitive Attributions

I'm now honing in on specific instances where human cognitive functions are directly or indirectly attributed to MLLMs. I've noted examples framing models as having "memory" and reacting to "strangeness," while others present them as evaluators, judges, analysts, and speculators. I'm carefully dissecting the language used to assess the level of directness and qualification in these attributions, and their implications for how we perceive model agency.

Investigating Model Attitudes

I am now closely examining an instance where GPT-4o is described as experiencing cognitive dissonance and altering its attitude. The focus is now on the mapping of human cognition to model behavior, exploring how the language projects subjective tension and epistemic agency onto the model. I am dissecting source and target domains (cognitive dissonance to attention weights) while acknowledging that there are concealed processes (the lack of subjective experience). I have identified hidden actors and am considering the implications of this direct, unacknowledged framing.

Dissecting Linguistic Markers

I'm now zeroing in on linguistic markers of confidence. I'm unpacking why LLMs generate highly probable language, even when the underlying content is factually incorrect. I'm focusing on the reframing of overconfidence as probable language. This includes an analysis of source and target domains (cognitive dissonance to attention weights), while acknowledging that there are concealed processes (the lack of subjective experience). I'm also examining the implications of this direct, unacknowledged framing. I'm carefully reviewing the provided examples, noting the shifts in emphasis and the potential implications of each reframing.

Revising Human Cognitive Attributes

I'm currently revising the anthropomorphic attributions. I'm replacing subjective descriptions, such as "speculating" and "changed its attitude," with explanations of LLM outputs and their underlying mechanisms. I am looking at how probabilistic associations, in-context learning, and visual feature extraction are more accurate descriptors, focusing on how these processes shape the model behavior, as well as considering how pareidolia and other similar constructs can be characterized within this framework.

Reframing Attributions & Outputs

I am now focusing on refining my analysis. I am replacing potentially subjective descriptors with more concrete explanations of LLM behavior. I'm focusing on the model's capacity to score similarly distributed inputs higher, which is characterized by distributional similarity. Also, I am now focusing on how LLMs generate outputs to statistically mirror human deviations in reasoning. Lastly, I am looking at how models emit outputs that have positive sentiment tokens in response to a prompt with positive framing.

Clarifying Metaphors & Agency

I'm now clarifying the problematic metaphors used, especially regarding agency. The transition from describing models as systems to autonomous entities using psychological terms is rapid. This shifts the focus from their training to the illusion of a flawed, human-like mind. This builds trust by treating the model as a psychological subject. I am also working on naming the corporations involved (OpenAI, Google, etc.) to clarify the origins of the models.

Analyzing Anthropomorphic Intensification

I'm now zeroing in on the escalation of anthropomorphism, contrasting early cautious descriptions with the intensified language used in empirical analysis. I'm noting the attribution of speculation to Claude, contrasting this with a hypothesis about a different memory mechanism in the introductory section. I'm also observing the systematic placement of responsibility on the "model" for biases, rather than on the corporations involved.

Mapping Discourse Frameworks

I'm now mapping the discourse frameworks, contrasting psychological terminology with mechanistic descriptions. My current focus is on the reframing: restoring developer agency and mechanistic explanations for model behaviors. I am checking the JSON schema, verifying the acknowledgment status, agency levels, mechanism basis, and actor visibility. The enums in the JSON schema are all validated, with a deep dive into the 'epistemic/conscious state' for greater precision.

Categorizing Epistemic States

I am now categorizing epistemic states, considering hedged vs. direct attributions to the models. I'm focusing on discrepancies, such as the initial acknowledgement of fundamental differences in memory mechanisms while simultaneously using anthropomorphic verbs like "speculating" and "reacting." The lack of company names for agency assessment is a central focus; I'll explain the need to restore accountability. I am focusing on refining the JSON schema enums. The framework must declare that "commercial incentives..." in 5.3.

Categorizing Explanation Types

I'm now diving into the explanation types used across the identified passages. I'm focusing on the primary and secondary explanation types assigned to each instance, specifically the intentional, reason-based, genetic, dispositional, and other identified types. I'm mapping these to the underlying claims about LLM behavior and considering whether these explanations align with mechanistic understanding or anthropomorphic framing. I am paying close attention to any inconsistencies.

Analysis Provenance

Run ID: 2026-10-15-a-multimedia-perspective-on-the-study-of-metaphor-nkmo3v Raw JSON: 2026-10-15-a-multimedia-perspective-on-the-study-of-metaphor-nkmo3v.json Framework: Metaphor Analysis v6.8 Schema Version: 3.4.3 Generated: 2026-10-15T10:11:20.035Z

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