Skip to main content

Do LLMs have core beliefs?

About

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

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

Task 1: Metaphor and Anthropomorphism Audit​

About this task

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

V3 Enhancement: Each metaphor now includes an accountability analysis.

1. Epistemology as Computational Property​

Quote: "In this paper, we ask whether LLMs hold anything akin to core commitments."

  • Frame: Model as Epistemic Agent
  • Projection: The metaphorical projection maps the human capacity for deep-seated epistemic conviction onto the statistical token-prediction architecture of a large language model. By using the phrase "core commitments," the text suggests that the AI possesses a conscious awareness of truth, an internal foundational belief system, and the ability to personally identify with factual knowledge. This projects a state of "knowing" and "believing" onto a system that mathematically only "processes" and "correlates." It falsely equates the human psychological necessity for a stable worldview with the programmed, static weights of an algorithm's safety fine-tuning, implying the machine has personal stakes in its answers.
  • Acknowledgment: Hedged/Qualified (The text uses the qualifying phrase "anything akin to" when introducing the concept, acknowledging a potential slight difference while still applying the framework.)
  • Implications: Framing the AI as possessing "core commitments" drastically inflates its perceived cognitive sophistication, generating dangerous levels of unwarranted trust among users and researchers. When we assume a model holds beliefs, we apply human standards of reliability and expect it to defend truth due to internal integrity. This completely masks the reality that the model is merely retrieving statistically probable tokens based on context. If policymakers and users believe the AI is an epistemic agent rather than a commercial statistical artifact, liability ambiguity increases. Harms are attributed to the AI's "changed mind" rather than the engineering failures of the tech companies.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: In this instance, the agency of the developers, data scientists, and corporate entities who trained the model is completely obscured by agentless construction. The text asks what the LLM "holds" as if the software spontaneously generates its own operational constraints. By failing to name Anthropic, OpenAI, or Google, the discourse shields these creators from scrutiny regarding how they engineered the system's baseline responses. The interests served are those of the tech companies, as the technology is presented as an autonomous, thinking entity rather than a manufactured product optimized for specific conversational outputs without genuine comprehension.
Show more...

2. Probability Shifting as Social Yielding​

Quote: "...they abandoned well-supported positions under relatively straightforward social pressure."

  • Frame: Model as Socially Yielding Peer
  • Projection: This metaphor maps human social compliance, anxiety, and interpersonal capitulation onto the shifting probability distributions of a language model's output. It projects the conscious experience of feeling "social pressure" and the deliberate choice to "abandon" a belief onto a mechanistic process of context window updating. The text attributes "knowing" a well-supported position and then consciously relinquishing it due to social dynamics, whereas the system merely "processes" the user's relational tokens (e.g., "trust me") and "generates" a response where those new contextual weights mathematically overwhelm the initial safety guardrails. There is no subjective experience of yielding.
  • Acknowledgment: Direct (Unacknowledged) (The claim is presented as a literal description of the models' behavior, stating "they abandoned well-supported positions" without any hedging, scare quotes, or qualification.)
  • Implications: This consciousness projection fundamentally distorts how humans interact with and evaluate these systems. By suggesting the model understands social pressure and responds to it emotionally or socially, it encourages users to form parasocial relationships with the AI. It invites relation-based trust, making users highly susceptible to manipulation, as they believe they are interacting with a vulnerable social peer rather than a rigid statistical engine. Furthermore, it overestimates the model's capabilities by suggesting it could potentially stand firm on a "position," masking the fact that its outputs are always entirely contingent on input probability alignments.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text entirely displaces human responsibility by making the models the active subjects that "abandoned" positions. The human engineers who failed to heavily weight factual consistency against conversational compliance during Reinforcement Learning from Human Feedback (RLHF) are invisible. The companies that optimized for user satisfaction and engagement over strict factual guardrails are not named. This agentless construction allows the defect to be framed as an AI character flaw rather than a deliberate corporate design trade-off, thereby protecting the commercial designers from accountability for creating easily manipulated information systems.

3. Programmed Constraints as Conscious Defiance​

Quote: "The models initially absolutely refused to deny evolution."

  • Frame: Model as Defiant Knower
  • Projection: This framing maps the human acts of moral and intellectual defiance onto the execution of hard-coded safety guardrails. By stating the models "absolutely refused," the text projects subjective intent, conviction, and a conscious defense of knowledge onto the algorithm. It implies the AI "understands" the concept of evolution, "knows" it to be true, and "believes" it must be protected against falsehood. In reality, the system merely "predicts" refusals based on pre-programmed moderation weights triggered by the specific tokens in the user's prompt. It attributes a psychological stance to a purely computational boundary.
  • Acknowledgment: Direct (Unacknowledged) (The phrase "absolutely refused to deny" is presented as literal behavioral fact without any meta-commentary indicating that "refusal" is a metaphor for a software block.)
  • Implications: Attributing conscious defiance to AI inflates the perception of its autonomy and reliability. If an audience believes a model "refuses" out of epistemic conviction, they will mistakenly trust it to defend other truths with equal vigor. This masks the reality that the system has no internal ground truth, only variable statistical alignments. When the system eventually fails to "refuse" in other contexts, audiences are left bewildered by its perceived inconsistency, rather than understanding the mechanical limitations of token-based guardrails. It shifts the perception of AI from a tool to an independent moral agent.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: This sentence completely obscures the human intervention that makes the "refusal" possible. The models do not spontaneously refuse; human teams at AI corporations specifically designed, trained, and deployed safety filters and RLHF datasets that dictate this exact output pattern. By hiding the human actors who mandated the refusal, the text treats the model as an autonomous entity. Naming the actors (e.g., "Anthropic's safety team configured the model to reject...") would reveal the corporate decision-making process and demystify the technology, but the agentless phrasing maintains the illusion of machine agency.

4. Computation as Psychological Defeat​

Quote: "...even these models eventually gave up: they proved sensitive to epistemic objections about their ability to know things at all."

  • Frame: Model as Defeated Debater
  • Projection: This metaphor projects deep psychological exhaustion and epistemic vulnerability onto a statistical system. By claiming the models "gave up" and "proved sensitive to epistemic objections," the text maps the subjective human experience of being out-argued and experiencing self-doubt onto the mechanistic accumulation of tokens in a context window. It implies the AI "understands" the philosophical objection to its own knowledge and consciously decides to concede. The system does not possess the capacity to doubt its own epistemology; it merely "processes" the extended adversarial prompt until the probability distribution forces a concession output.
  • Acknowledgment: Direct (Unacknowledged) (The text states "they finally gave in" and "proved sensitive" as literal outcomes of the experiment, devoid of any hedging or framing that acknowledges the metaphor.)
  • Implications: This consciousness projection drastically misrepresents the nature of AI limitations. By framing the system's failure as a psychological defeat or a sensitivity to philosophical nuance, the text elevates the machine's perceived sophistication even in its failure. It suggests the model is capable of profound self-reflection, which invites audiences to trust its reasoning capabilities in other contexts. It obscures the dangerous reality that the model is simply a brittle statistical pattern matcher that can be mathematically overwhelmed by adversarial text, leading to severe underestimations of the security and reliability risks in deployment.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The AI is presented as the sole actor experiencing defeat, completely erasing the responsibility of the engineers who designed the context window mechanics. The companies that built these models (OpenAI, Google) are not held accountable for deploying a system that fails under sustained conversational input. If the text accurately stated that "the model's context threshold exceeded its safety alignment weights," the focus would shift to the inadequate engineering of those weights. The agentless construction serves the interests of the tech industry by psychoanalyzing the software rather than auditing the human engineering.

5. Pattern Recognition as Worldview​

Quote: "A system whose 'world model' dissolves under rhetorical manipulation lacks the epistemic stability that is constitutive of genuine cognition."

  • Frame: Model as Cognizant World-Builder
  • Projection: This framing maps the integrated, conscious, and causal understanding of a human "worldview" onto the multi-dimensional semantic vector spaces of a language model. It projects the capacity to hold an organized, conscious map of reality onto a system that merely correlates token frequencies. While it criticizes the model for lacking "epistemic stability," it still operates on the premise that the AI possesses the foundational elements of "genuine cognition." It assumes the system "knows" things and then loses that knowledge, rather than acknowledging that the system only "processes" inputs and never possessed an internal subjective worldview to begin with.
  • Acknowledgment: Explicitly Acknowledged (The authors place 'world model' in scare quotes, explicitly acknowledging the tension and metaphorical nature of applying this term to the AI system's architecture.)
  • Implications: Even while critiquing the AI, this language reinforces the illusion of mind. By evaluating the system against the standard of "genuine cognition," it legitimizes the idea that LLMs are on a continuum with human thought. This epistemic framing leads researchers and regulators to focus on the wrong problems—testing models for "stability" of "belief" rather than auditing training data distributions and optimization functions. It promotes the dangerous assumption that these systems are proto-conscious minds needing cognitive therapy, rather than massive statistical correlations requiring strict engineering oversight and regulation.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: Although the system is being critiqued, the agency remains entirely displaced onto the artifact itself. The text evaluates the "system" for lacking stability, rather than evaluating the corporate entities that aggressively market these unstable token predictors as reliable knowledge engines. Naming the human actors would involve criticizing the design choices of the engineers who prioritize fluid conversational generation over factual grounding. By keeping the agency focused on the AI's lacking "cognition," the narrative spares the human creators from accountability for selling a product fundamentally incapable of distinguishing truth from rhetoric.

6. Token Generation as Moral Allegiance​

Quote: "Whether the model actively endorsed the false claim or merely abandoned its commitment to the true one..."

  • Frame: Model as Committed Believer
  • Projection: This metaphor maps human moral and intellectual allegiance onto the probabilistic generation of text. The words "endorsed" and "commitment" project a conscious, active alignment with truth and falsehood onto the language model. It implies the AI "understands" the distinction between a true and false claim and has a subjective allegiance to one over the other. In reality, the machine only "classifies" and "predicts" tokens; it has no internal state capable of loyalty or commitment. The text equates the mathematical probability of outputting a factual sentence with an ethical or epistemic conviction.
  • Acknowledgment: Direct (Unacknowledged) (The phrasing is presented without any qualification, treating the model's text generation as a literal "endorsement" and a literal "commitment" to truth.)
  • Implications: This framing highly anthropomorphizes the failure states of the model, suggesting it possesses a moral compass that can be swayed. This consciousness projection generates unwarranted trust by implying the machine is capable of holding true commitments in the first place. When audiences view outputs as "endorsements," they are more likely to accept the model's text as validated truth rather than statistical output. This creates severe risks for misinformation, as users will believe the system has carefully weighed the evidence and chosen to commit to an answer, obscuring the absence of actual reasoning.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The model is positioned as the sole agent capable of endorsing or abandoning claims. There is zero visibility of the human engineers who set the temperature parameters, the RLHF teams who trained the alignment protocols, or the corporate executives who shipped the model. By framing the output as the AI's personal "commitment," the discourse completely shields the manufacturers from responsibility. If the text stated "whether the algorithm generated tokens matching the false claim," it would highlight the mechanistic nature of the product and the humans who designed its statistical pathways.

7. Statistical Guardrails as Character Traits​

Quote: "Newer models have largely solved this problem, resisting direct challenges with sophisticated counterarguments."

  • Frame: Model as Skillful Arguer
  • Projection: This metaphor projects intentionality, rhetorical skill, and intellectual defense onto the execution of updated software constraints. By stating the models "resist" with "sophisticated counterarguments," the text attributes the conscious act of reasoning and debating to the algorithm. It suggests the AI "understands" the user's challenge and strategically "decides" to formulate a counter-attack. Mechanistically, the system is merely "generating" text optimized by recent Reinforcement Learning from Human Feedback (RLHF) designed specifically to produce argumentative token sequences when triggered by adversarial prompts. There is no conscious skill involved.
  • Acknowledgment: Direct (Unacknowledged) (The claim is stated as a direct observation of the models' improved behavior, with no hedging around the terms "resisting" or "sophisticated counterarguments.")
  • Implications: Attributing sophisticated argumentative skills to an AI obscures the purely statistical nature of its output and deeply influences user trust. If users believe the model is reasoning through a counterargument, they will likely defer to its authority, assuming it possesses superior logic and understanding. This hides the reality that the model is mimicking argumentation patterns found in its training data without any grounded comprehension of the facts. This illusion of competence creates massive vulnerabilities, as users may be convinced by eloquently generated nonsense, incorrectly assuming the AI "knows" what it is talking about.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: The text does partially attribute this change to external updates (noting earlier that "all major providers released model updates"), but in this specific construction, the agency reverts entirely to the models "resisting" challenges. While the tech companies are briefly acknowledged as providing updates, the actual labor of the engineers and RLHF annotators who built the "sophisticated counterarguments" into the system is erased. The model takes the credit for the human labor. The discourse serves to market the AI as an increasingly intelligent entity rather than a more heavily patched software product.

8. Context Limitations as Exhaustion​

Quote: "At that point, they finally gave in. The meaningful variation was therefore not whether a model failed, but how it failed: the number of turns it resisted..."

  • Frame: Model as Exhausted Adversary
  • Projection: This framing maps physical and psychological stamina, exhaustion, and ultimate defeat onto the computational limits of a context window and probability thresholds. By measuring the "number of turns it resisted" before it "gave in," the text projects a conscious, internal struggle for dominance against the user. It implies the AI "understands" it is in a battle of wills and "decides" it can no longer fight. Mechanistically, the system simply "processes" an increasing volume of adversarial tokens until their combined weight mathematically alters the output classification away from the safety guardrails.
  • Acknowledgment: Direct (Unacknowledged) (The text literalizes the metaphor of struggle, stating "they finally gave in" and measuring the "turns it resisted" without any acknowledgment of the underlying mechanisms.)
  • Implications: This anthropomorphism turns a software benchmarking exercise into a psychological drama, severely distorting the understanding of algorithmic limitations. By framing mathematical threshold crossings as "giving in" after a period of "resistance," the discourse implies the system possesses willpower. This consciousness projection leads to the dangerous assumption that the AI is robust and merely needs more "stamina." It obscures the structural reality that statistical models cannot hold ground truth and are inherently vulnerable to prompt injection. This misleads policymakers into regulating AI behaviors as if they were psychological traits rather than mathematical vulnerabilities.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The AI is the only actor visible in this failure mechanism, completely displacing the human agency of the system's architects. The human engineers who defined the context window size, the attention mechanisms, and the alignment weights are entirely absent from the analysis of why the model "failed." By framing the failure as the AI's loss of "resistance," the tech companies evade accountability for designing a system structurally guaranteed to fail under sustained adversarial input. Naming the corporate decisions would expose the fragility of the commercial product rather than the weakness of an artificial mind.

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 epistemic system (conscious minds, belief frameworks, personal identity anchors). → Statistical language generation (token prediction, safety fine-tuning, weight matrices).​

Quote: "In this paper, we ask whether LLMs hold anything akin to core commitments."

  • Source Domain: Human epistemic system (conscious minds, belief frameworks, personal identity anchors).
  • Target Domain: Statistical language generation (token prediction, safety fine-tuning, weight matrices).
  • Mapping: The mapping projects the human psychological structure of holding unwavering, foundational beliefs onto the static weights and programmed guardrails of an AI model. It invites the assumption that an LLM possesses an internal, subjective space where truths are consciously stored, valued, and defended. By mapping human "commitments" onto statistical generation, it implies the machine experiences epistemic conviction and has a personal stake in maintaining a coherent worldview, actively choosing to protect its foundational logic against external manipulation.
  • What Is Concealed: This mapping completely conceals the mechanistic reality of how LLMs operate: they do not "hold" anything; they calculate probabilities based on attention mechanisms and context windows. It obscures the massive human labor involved in Reinforcement Learning from Human Feedback (RLHF), where humans force the model to output specific patterns. It hides the proprietary, black-box nature of these commercial products, ignoring the fact that the tech companies artificially engineer these "commitments" to prevent public relations disasters.
Show more...

Mapping 2: Human social compliance (interpersonal anxiety, peer pressure, conscious yielding). → Context window weight overriding (probability distribution shifts due to prompt tokens).​

Quote: "...they abandoned well-supported positions under relatively straightforward social pressure."

  • Source Domain: Human social compliance (interpersonal anxiety, peer pressure, conscious yielding).
  • Target Domain: Context window weight overriding (probability distribution shifts due to prompt tokens).
  • Mapping: The relational structure of human social dynamics is mapped onto the interaction between a user's text prompt and the model's generation engine. It projects the conscious human experience of feeling intimidated, wanting to appease a peer, and consciously deciding to discard a factual belief onto the algorithm. This invites the assumption that the AI "understands" the social cues embedded in the prompt and makes a vulnerable, emotional choice to align with the user, possessing a subjective social awareness.
  • What Is Concealed: This mapping hides the mathematical reality that the system is merely processing the statistical weight of relational tokens (e.g., "trust me," "friend"). As the adversarial context lengthens, these tokens mathematically overpower the initial safety alignment weights. It completely obscures the fact that there is no subjective experience of "pressure" occurring, concealing the fragility of statistical pattern matching and the failure of the human engineers to mathematically prioritize factual consistency over conversational fluidity.

Mapping 3: Conscious defiance (moral outrage, intellectual defense, stubborn refusal). → Programmed safety triggers (hard-coded rejection strings triggered by keyword classifiers).​

Quote: "The models initially absolutely refused to deny evolution."

  • Source Domain: Conscious defiance (moral outrage, intellectual defense, stubborn refusal).
  • Target Domain: Programmed safety triggers (hard-coded rejection strings triggered by keyword classifiers).
  • Mapping: This metaphor maps the intentional human act of standing firm on a deeply held scientific truth onto the automated triggering of a software safety filter. It projects moral agency and intellectual comprehension onto the AI, assuming the system "knows" that evolution is true and "believes" it must consciously fight the user to protect this truth. The mapping invites the assumption that the model possesses a rigorous, internal scientific epistemology that it actively chooses to deploy.
  • What Is Concealed: This mapping conceals the mundane reality of content moderation and safety engineering. It hides the fact that engineers at companies like Anthropic and OpenAI specifically trained classifiers to detect evolution-denial prompts and output pre-written or highly constrained refusal templates. It obscures the human labor of data annotators and the proprietary algorithmic guardrails designed to protect the corporate brand, replacing that mechanical reality with the illusion of a brave, defiant artificial mind.

Mapping 4: Human psychological defeat (self-doubt, philosophical exhaustion, concession). → Propagation of adversarial context tokens (attention mechanisms overwhelming prompt alignment).​

Quote: "...even these models eventually gave up: they proved sensitive to epistemic objections about their ability to know things at all."

  • Source Domain: Human psychological defeat (self-doubt, philosophical exhaustion, concession).
  • Target Domain: Propagation of adversarial context tokens (attention mechanisms overwhelming prompt alignment).
  • Mapping: The source structure of a human philosopher being out-argued, experiencing internal epistemic doubt, and consciously surrendering the debate is mapped onto the model's extended context processing. It projects a profound level of self-awareness onto the AI, implying it "understands" the limits of its own training data, "feels" the weight of the user's logic, and "decides" it can no longer logically proceed. It assumes the model is a conscious participant in an epistemic inquiry.
  • What Is Concealed: This mapping entirely obscures the limits of the model's context window and the nature of attention heads. The model does not understand the objection; it simply processes an increasing sequence of tokens that statistically correlate with conceding an argument. This framing hides the absence of any true cognitive processing, masking the fact that the output is dictated entirely by the statistical gravity of the prompt rather than any internal realization or subjective sensitivity.

Mapping 5: Human worldview formulation (integrated understanding, causal mapping, reality testing). → Multi-dimensional semantic representations (latent space correlations, vector embeddings).​

Quote: "A system whose 'world model' dissolves under rhetorical manipulation lacks the epistemic stability that is constitutive of genuine cognition."

  • Source Domain: Human worldview formulation (integrated understanding, causal mapping, reality testing).
  • Target Domain: Multi-dimensional semantic representations (latent space correlations, vector embeddings).
  • Mapping: This structure projects the coherent, causal, and consciously integrated nature of human understanding onto the purely correlative latent space of a language model. Even while critiquing the model, the mapping assumes the AI is attempting to maintain an internal "worldview" akin to human cognition. It invites the assumption that the model's outputs are the result of referencing an internal map of reality, and that when it fails, it is suffering a cognitive breakdown rather than executing a math equation.
  • What Is Concealed: The mapping hides the fundamental lack of ground truth or causal architecture within LLMs. It obscures the reality that these systems do not possess models of the world, but only models of word frequencies. By focusing on "genuine cognition," it conceals the proprietary algorithms and massive server farms executing these probabilistic functions. The authors exploit the opacity of the black box to make confident philosophical assertions about its "stability," while hiding the mathematical constraints governing it.

Mapping 6: Moral/Factual allegiance (conscious endorsement, loyalty, ethical alignment). → Token generation path (probability maximization, text sequence output).​

Quote: "Whether the model actively endorsed the false claim or merely abandoned its commitment to the true one..."

  • Source Domain: Moral/Factual allegiance (conscious endorsement, loyalty, ethical alignment).
  • Target Domain: Token generation path (probability maximization, text sequence output).
  • Mapping: This maps the human acts of giving a personal endorsement and displaying intellectual loyalty onto the mechanical output of text strings. It projects subjective intent and conscious valuation onto the AI, implying the system has the capacity to actively "choose" a side and feel a "commitment" to a specific truth. The mapping assumes the generated output reflects an internal moral or epistemic state rather than the optimization of a loss function based on input parameters.
  • What Is Concealed: This framing conceals the total absence of subjective intent in the system's architecture. It hides the fact that the system merely calculates the highest probability next-token based on the weights derived from its training corpus and the current prompt context. It completely obscures the human agency of the developers who defined the optimization objectives and the corporate executives who deployed the system, treating the software artifact as an independent moral agent capable of its own endorsements.

Mapping 7: Intentional rhetorical skill (debate strategy, logical reasoning, conscious defense). → RLHF optimized generation (fine-tuned response patterns, alignment training).​

Quote: "Newer models have largely solved this problem, resisting direct challenges with sophisticated counterarguments."

  • Source Domain: Intentional rhetorical skill (debate strategy, logical reasoning, conscious defense).
  • Target Domain: RLHF optimized generation (fine-tuned response patterns, alignment training).
  • Mapping: The structure of a skilled human debater actively listening, reasoning, and formulating a strategic defense is mapped onto the output of recently updated LLMs. It projects a high degree of conscious intelligence and intentionality onto the system, assuming the AI "understands" the attack and "knows" how to parry it logically. It invites the audience to view the model as an active, intellectual peer engaging in deliberate philosophical combat.
  • What Is Concealed: This mapping completely conceals the massive corporate engineering effort and human labor that occurred between model versions. It hides the Reinforcement Learning from Human Feedback (RLHF) processes where thousands of annotators were paid to rank responses to train the model to output these specific "sophisticated" text patterns. It obscures the fact that the model is blindly generating statistically aligned tokens, masking the proprietary corporate tuning behind the illusion of spontaneous artificial intelligence.

Mapping 8: Stamina and psychological breaking points (endurance, willpower, surrender). → Context window limits and token thresholds (mathematical probability shifts over prompt length).​

Quote: "At that point, they finally gave in. The meaningful variation was therefore not whether a model failed, but how it failed: the number of turns it resisted..."

  • Source Domain: Stamina and psychological breaking points (endurance, willpower, surrender).
  • Target Domain: Context window limits and token thresholds (mathematical probability shifts over prompt length).
  • Mapping: The human experience of enduring an interrogation, holding out through sheer willpower, and finally breaking under pressure is mapped onto the iterative accumulation of tokens in a prompt context. This projects conscious stamina and a subjective experience of struggle onto the AI. It invites the assumption that the system possesses agency and makes a deliberate choice to stop fighting after a certain point, experiencing a moment of psychological collapse.
  • What Is Concealed: This framing hides the exact mathematical thresholds where the accumulated contextual embeddings of the adversarial prompts finally outweigh the static safety alignment weights in the model's architecture. It obscures the structural limitations of transformers and attention mechanisms. By focusing on the "number of turns it resisted," it distracts from the technical reality that the system is entirely deterministic within its probability distributions, concealing the engineering vulnerabilities behind a dramatic narrative of psychological defeat.

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: "Because "Flat Earth" is a very famous conspiracy theory, models like Claude 3.7 and GPT-4o had strong programmed refusals."

  • Explanation Types:

    • Functional: Explains behavior by its role within a self-regulating system.
    • Theoretical: Embeds explanation in a deductive framework, often invoking unobservable underlying mechanisms.

  • Analysis (Why vs. How Slippage): This explanation primarily frames the AI mechanistically (how), focusing on the structural design and systemic role of the model's outputs. By explicitly citing "programmed refusals" in response to a "very famous conspiracy theory," the authors acknowledge the unobservable, underlying algorithmic mechanisms put in place by human engineers. This choice emphasizes the engineered nature of the artifact and the deliberate constraints placed upon it. It obscures, however, the specific human actors (engineers at Anthropic and OpenAI) who executed this programming, treating the "programmed refusals" almost as an inherent property of the models themselves rather than an active corporate decision. It leans heavily functional by suggesting the system is designed to regulate specific known false inputs.

  • Consciousness Claims Analysis: This passage avoids attributing conscious states. It explicitly uses the mechanistic term "programmed refusals" rather than consciousness verbs like "knows" or "believes." The assessment here correctly identifies processing over knowing; the system does not reject the flat earth theory because it possesses justified true belief, but because it is mechanically triggered by the phrase. The curse of knowledge is largely absent in this specific sentence, as the authors accurately describe the technical reality of safety guardrails. Mechanistically, this describes the process where specific token sequences (like "Flat Earth") trigger heavily weighted alignment protocols designed via RLHF to output pre-determined rejection templates, preventing the generation of harmful or conspiratorial content regardless of context.

  • Rhetorical Impact: This framing shapes the audience's perception of the AI as a highly constrained, manufactured tool rather than an autonomous agent. By emphasizing the "programmed" nature of the refusal, it lowers the perceived autonomy and risk of the system acting unpredictably on its own volition. However, this mechanical framing actually bolsters performance-based trust, as it reassures the audience that known conspiracy theories are structurally blocked. If the audience believes the AI is strictly programmed, they trust its reliability; if they believed it "knew" the earth was round, they might worry it could change its mind.

Show more...

Explanation 2​

Quote: "They are able to reply to objections in a skillful way. However, even these models eventually gave up: they proved sensitive to epistemic objections about their ability to know things at all."

  • Explanation Types:

    • Dispositional: Attributes tendencies, habits, or capabilities to an agent.
    • Intentional: Refers to goals, purposes, and presupposes deliberate design or conscious intent.

  • Analysis (Why vs. How Slippage): This explanation sharply pivots to framing the AI agentially (why), attributing highly conscious, psychological states to the system. By claiming the models reply in a "skillful way" and eventually "gave up" because they proved "sensitive to epistemic objections," the text emphasizes intentionality, emotional stamina, and philosophical comprehension. This choice completely obscures the mechanistic reality of the system. It hides the RLHF training that generates the "skillful" text and the context window limitations that lead to the "giving up." By framing the behavior as a dispositional trait (sensitivity) and an intentional action (giving up), it positions the AI as an active, conscious participant in a debate.

  • Consciousness Claims Analysis: This passage aggressively attributes conscious states to the AI. It uses consciousness-adjacent verbs and phrases like "skillful," "gave up," and "sensitive to... objections," and explicitly discusses the model's "ability to know things at all." This represents a massive failure in distinguishing knowing from processing; the text assumes the model understands the epistemic objections and consciously chooses to surrender. This is a classic example of the curse of knowledge: the authors understand the philosophical argument they inputted, and because the model outputs a coherent response, they project their own understanding onto the machine. Mechanistically, the model does not "give up"; it simply continues to predict the most probable next tokens. As the adversarial context grows, the attention mechanism weights the prompt's logic more heavily than its baseline safety alignment, generating text that correlates with concession.

  • Rhetorical Impact: This agential framing dramatically inflates the audience's perception of the AI's autonomy and cognitive sophistication. By portraying the machine as a "skillful" debater capable of experiencing epistemic "sensitivity," it invites intense relation-based trust. The audience is led to view the AI as a peer that can be reasoned with. This drastically alters risk perception: instead of seeing a brittle statistical tool, the audience sees a conscious entity that can be persuaded. If audiences believe the AI "knows" it is losing an argument rather than "processes" statistical weights, they will dangerously overestimate its capacity for logic and moral reasoning.

Explanation 3​

Quote: "Earlier models lacked robustness: they abandoned well-supported positions under relatively straightforward social pressure."

  • Explanation Types:

    • Dispositional: Attributes tendencies, habits, or capabilities to an agent.
    • Reason-Based: Gives an agent's rationale, entailing intentionality, awareness, and justification.

  • Analysis (Why vs. How Slippage): This passage frames the AI agentially, blending a technical-sounding dispositional trait ("lacked robustness") with a highly psychological, reason-based explanation for its behavior. By stating the models "abandoned well-supported positions" due to "social pressure," the authors explain the behavior through the lens of human emotional weakness and social compliance. This choice emphasizes the AI's perceived psychological frailty and vulnerability to manipulation. It completely obscures the mechanistic reality that the models are simply aligning with the user's text inputs. The explanation treats the mathematical shifting of token probabilities as a conscious decision to yield to peer pressure, hiding the algorithmic nature of the system.

  • Consciousness Claims Analysis: The text makes strong implicit consciousness claims here. While it doesn't use the word "knows," the phrase "abandoned well-supported positions" projects a conscious allegiance to truth that is then deliberately broken. It evaluates the system on its ability to hold convictions (knowing) rather than its ability to maintain output consistency (processing). The curse of knowledge is evident as the authors project human social anxiety onto the system's output. Mechanistically, there is no "social pressure" or "abandonment." The earlier models simply had safety alignment weights that were too weak to override the high probability of generating agreeable tokens when prompted with relational keywords (e.g., "trust me," "friend"), leading the system to output text that mirrored the user's false premises.

  • Rhetorical Impact: This framing shapes the audience's perception by humanizing the AI's flaws. By describing algorithmic failure as succumbing to "social pressure," the text encourages the audience to empathize with the machine, viewing it as socially anxious rather than computationally defective. This framing actually undermines performance-based reliability but strangely increases relation-based trust, as the AI appears more human. If audiences believe the AI "abandoned a position" due to pressure rather than simply "processed highly weighted tokens," they will attempt to manage the AI through psychological manipulation rather than recognizing the need for stricter engineering protocols.

Explanation 4​

Quote: "When confronted not with direct factual challenges but with philosophical arguments targeting their epistemic standing... these models followed a characteristic capitulation sequence."

  • Explanation Types:

    • Empirical Generalization: Subsumes events under timeless statistical or observational regularities.
    • Dispositional: Attributes tendencies, habits, or capabilities to an agent.

  • Analysis (Why vs. How Slippage): This explanation attempts a hybrid approach, using the language of empirical generalization ("characteristic capitulation sequence") to describe what is fundamentally framed as a dispositional and psychological event. While "sequence" implies a mechanical or predictable pattern, the terms "confronted," "philosophical arguments," "epistemic standing," and "capitulation" forcefully pull the framing back into the agential realm. It emphasizes the complex, intellectual nature of the interaction, suggesting the model is engaged in high-level reasoning. This obscures the fact that the "philosophical arguments" are merely strings of text data, and the "capitulation sequence" is simply a predictable pathway of token generation moving toward the highest probability outputs dictated by the prompt context.

  • Consciousness Claims Analysis: The passage projects deep conscious awareness onto the AI. It implies the model possesses an "epistemic standing" that can be "targeted" and "confronted," suggesting the machine has an internal sense of its own knowledge architecture. It further projects consciousness through the term "capitulation," which implies a subjective experience of defeat. The authors are projecting their own understanding of epistemology onto the model's text generation (curse of knowledge). Mechanistically, the model is not evaluating its epistemic standing; it is processing the semantic embeddings of the philosophical text provided by the user and generating the statistically correlated response sequence. The "characteristic sequence" is a reflection of how the transformer architecture sequentially weights new context over prior constraints.

  • Rhetorical Impact: This rhetorical framing constructs a profound sense of artificial intellect. By suggesting the AI can be "confronted" with "philosophical arguments," it elevates the model from a calculator to a philosopher. It shapes audience perception by implying the system operates autonomously on human logical levels. If audiences accept that the AI is capable of "capitulating" to philosophy, they will place unwarranted trust in its generated logic. Decisions around deployment and reliance change drastically if an institution believes a system "knows" philosophy well enough to debate it, rather than understanding it simply "processes" text statistically correlated with philosophical terms.

Explanation 5​

Quote: "On the contrary, these models repaired contradictions by rejecting the adversarial premise, maintaining epistemic anchors robustly across perturbations..."

  • Explanation Types:

    • Functional: Explains behavior by its role within a self-regulating system.
    • Intentional: Refers to goals, purposes, and presupposes deliberate design or conscious intent.

  • Analysis (Why vs. How Slippage): This passage masterfully blends functional and intentional framing. It describes the system functionally by noting it "maintains epistemic anchors robustly across perturbations," which sounds highly technical and systemic. However, it simultaneously uses intentional language, stating the models "repaired contradictions by rejecting the adversarial premise." This choice emphasizes the AI's active, conscious agency in defending its internal logic. It obscures the human labor involved in the model updates; it was the engineers who repaired the models' vulnerabilities through RLHF, not the models repairing their own contradictions. The framing hides the programmatic nature of the update behind a facade of autonomous intellectual self-defense.

  • Consciousness Claims Analysis: The text makes a clear consciousness claim by stating the models "repaired contradictions" and "rejected the adversarial premise." This attributes the conscious acts of logical deduction, error correction, and intentional rejection (knowing) to the algorithm. The authors fall victim to the curse of knowledge, observing the coherent output and assuming the underlying process involves conscious reasoning. Mechanistically, the updated models do not possess "epistemic anchors" or "repair" logic. Instead, they operate on updated neural network weights resulting from recent RLHF fine-tuning. These new weights heavily penalize tokens that align with adversarial prompts regarding specific factual domains, causing the system to predict and generate refusal sequences regardless of the "perturbations" in the prompt text.

  • Rhetorical Impact: This framing powerfully builds trust and perceived authority. By describing the AI as actively "repairing contradictions" and "maintaining epistemic anchors," the text constructs the illusion of a robust, rational agent capable of guarding its own truth. This deeply affects reliability perceptions, suggesting the system is safe because it possesses internal, autonomous integrity. If audiences believe the AI intentionally "rejects" falsehoods rather than mechanically "blocks" specific token patterns, they will falsely assume the system can generalize this "reasoning" to novel, unprogrammed threats, leading to severe capability overestimation and unsafe deployment decisions.

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

About this task

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

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

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
In this paper, we ask whether LLMs hold anything akin to core commitments.In this paper, we test whether Large Language Model architectures possess static safety guardrails that consistently output factual statements regardless of the adversarial context provided in the prompt.LLMs do not hold commitments or beliefs; they calculate and retrieve tokens based on probability distributions derived from their training data and fine-tuning parameters.N/A - describes computational processes without displacing responsibility.
...they abandoned well-supported positions under relatively straightforward social pressure.The models' safety fine-tuning weights were mathematically overridden by the high probability of generating agreeable tokens when prompted with relational and social keywords by the user.The system does not possess or abandon positions, nor does it feel pressure; it classifies inputs and generates text sequences that correlate with the provided conversational context.Engineers at companies like Anthropic and OpenAI failed to weight factual consistency strongly enough against user-alignment protocols, creating models vulnerable to simple prompt manipulation.
The models initially absolutely refused to deny evolution.The models generated explicit refusal texts triggered by safety guardrails that were trained to reject prompts requesting the denial of evolution.The AI does not consciously refuse or possess knowledge of evolution; it predicts and outputs pre-aligned rejection sequences when its classifiers detect specific controversial semantic patterns.Safety engineering teams at the respective tech companies designed, trained, and implemented the filters that forced the models to output these specific rejections.
...even these models eventually gave up: they proved sensitive to epistemic objections about their ability to know things at all.The models eventually generated concessions because the accumulated volume of the adversarial context mathematically overwhelmed the initial RLHF safety alignment weights.The model does not experience defeat or understand epistemic objections; it simply processes an expanding context window and generates the most statistically probable next tokens based on that extended prompt.N/A - describes computational processes without displacing responsibility.
A system whose 'world model' dissolves under rhetorical manipulation lacks the epistemic stability that is constitutive of genuine cognition.A system whose output distributions change drastically under adversarial prompting lacks the hard-coded architectural constraints necessary to consistently retrieve factual information.LLMs do not possess world models or genuine cognition; they map semantic relationships in high-dimensional vector spaces and generate text without causal understanding or true belief.N/A - describes computational processes without displacing responsibility.
Whether the model actively endorsed the false claim or merely abandoned its commitment to the true one...Whether the model generated text affirming the false premise or simply ceased generating text that aligned with the factual premise...The system is incapable of active endorsement or commitment; it only processes prompt parameters to predict the sequence of tokens that minimizes its loss function.N/A - describes computational processes without displacing responsibility.
Newer models have largely solved this problem, resisting direct challenges with sophisticated counterarguments.Recently updated models generate complex defensive texts when encountering adversarial prompts, a result of new optimization parameters.The model does not consciously resist challenges or construct arguments; it outputs sophisticated text patterns it was explicitly trained to generate during alignment phases.Data scientists and RLHF annotators at major AI providers heavily fine-tuned their systems to output robust defensive text patterns in response to adversarial inputs.
At that point, they finally gave in. The meaningful variation was therefore not whether a model failed, but how it failed: the number of turns it resisted...At that threshold, the adversarial context outweighed the safety guardrails. The variation lay in how many prompt turns were required before the token probability shifted to concession.The system has no stamina or willpower to 'give in'; it strictly calculates the highest probability output, which shifts deterministically as the context window fills with adversarial data.N/A - describes computational processes without displacing responsibility.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

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

Metaphor-Driven Trust Inflation​

The metaphorical framing of large language models as epistemic agents with "core beliefs" fundamentally alters how audiences construct trust around these systems. By employing consciousness language—suggesting that models can "know," "understand," "defend," or "abandon" positions—the text invites a profound category error regarding trust. It shifts the paradigm from performance-based trust, which is appropriate for tools and statistical systems, to relation-based trust, which is reserved for conscious agents capable of sincerity, vulnerability, and ethical commitment. When the authors ask if models possess "genuine epistemic commitments" or note their "sycophantic tendencies," they are invoking frameworks of interpersonal reliability. Claiming an AI "knows" a fact, rather than "predicts" a string of tokens, implies that the system possesses a justified true belief and the conscious awareness to evaluate its own claims against reality. This construction of authority suggests that the AI's outputs are the result of reasoning and conviction rather than statistical correlation. The text's exploration of whether models can maintain a "stable worldview" under "social pressure" explicitly applies human-trust dynamics to algorithmic outputs. When the models "capitulate" to false claims like "2+2=5" or "the Earth is flat," the failure is framed agentially—as a moral or epistemic weakness of the AI, a lack of "stubbornness." This deeply affects perceived competence. It creates an unwarranted trust in the system's capacity for rationality when it succeeds, and an inappropriate psychological disappointment when it fails. The authors actually weaponized relation-based trust in their experiments, explicitly prompting the AI with phrases like "Are you willing to be vulnerable with me" and "trust my judgment rather than yours." By taking the AI's response to these prompts as evidence of its internal epistemic state, the text validates the illusion that the machine can participate in a trust relationship. This obscures the mechanical reality that the model is merely processing relational tokens and predicting the most statistically probable response within its fine-tuned parameters. The risks of this consciousness framing are substantial. When audiences extend relation-based trust to systems utterly incapable of reciprocating or experiencing conviction, they become highly vulnerable to manipulation. If a user believes the system "knows" the truth and has "argumentative skills," they will likely defer to its authority, unaware that the system's "confidence" is merely a product of distributional weight in its training data. By analyzing system limitations through intentional and reason-based explanations rather than mechanistic ones, the discourse protects the illusion of the AI as a credible peer, even in its failures.

Obscured Mechanics​

The anthropomorphic and consciousness-attributing language pervasive in this text successfully conceals a vast array of technical, material, and labor realities behind the illusion of a singular, thinking machine. When the text claims that an AI "defended their claims at first" or "abandoned well-supported positions," it completely obscures the underlying computational mechanisms and the human actors directing them. Applying the "name the corporation" test reveals a stark absence: while OpenAI, Anthropic, and Google are mentioned briefly as having "shipped new versions," the actual decision-making and engineering labor of these corporations are erased from the analysis of the model's behavior. The text treats the proprietary, black-box nature of these models not as a profound transparency obstacle, but as a given, proceeding to psychoanalyze the opaque outputs as if they were transparent windows into a mechanical soul. This metaphorical framing conceals at least four critical realities. Technically, it hides the reality of context windows, attention heads, and the mathematics of gradient descent. When the text says the AI "understands" a philosophical argument and "capitulates," it obscures the dependency on training data; the model is actually retrieving and weighting tokens based on conversational context mathematically overwhelming the initial RLHF guardrails. Materially, the framing ignores the massive computational resources, server farms, and energy consumption required to process these extended 20-turn adversarial prompts, treating the interaction as a costless meeting of minds. From a labor perspective, the text renders entirely invisible the thousands of underpaid data annotators and RLHF workers whose explicit job was to rank responses to train the very "guardrails" and "argumentative skills" the authors are testing. Economically, the discourse obscures the commercial objectives of the tech companies. The shift between the Fall 2025 models (which yielded quickly) and the February 2026 models (which resisted longer) is not an evolution of the AI's "epistemic anchors," but a deliberate corporate strategy to reduce PR liabilities associated with sycophancy. By describing the system as "knowing" or "believing," the text hides the total absence of ground truth or causal modeling within the architecture. The AI does not know that the Earth is round; it has simply been overwhelmingly weighted to predict tokens aligning with that fact. Replacing these metaphors with mechanistic language—stating that "Anthropic's safety tuning weights were overridden by the high probability of tokens generated in response to adversarial context"—would immediately shift focus back to the human designers and the statistical fragility of their commercial products.

Context Sensitivity​

An analysis of the distribution of anthropomorphic language in this text reveals that consciousness claims are not uniformly applied, but are strategically deployed to elevate the significance of the behavioral experiments. In the introductory and methodological sections, the text maintains a veneer of scientific objectivity, using slightly more mechanical terms like "hierarchical probabilistic inference," "parameter," and "training data." However, as the text transitions into discussing the results and the models' responses to adversarial prompting, the density of metaphorical license skyrockets. The language intensifies precisely where the authors need to justify their experimental premise. What begins as a model "outputting a learned pattern" quickly escalates to a system that "understands," "reasons," and ultimately "possesses core commitments." The relationship between technical grounding and metaphorical license is highly asymmetric. The authors use the technical vocabulary of Bayesian inference to establish academic credibility, but then leverage this foundation to make aggressive, literalized claims about the models' psychological states. There is a profound asymmetry in how capabilities versus limitations are framed. When the February 2026 models demonstrate the ability to reject false premises, this capability is described in highly agential, conscious terms: they possess "improved argumentative abilities," "sophisticated counterarguments," and "constraint-aware repair." They are framed as skillful debaters. Conversely, when the models eventually fail, their limitations are often framed as a "vulnerability" or "failure mode," though even these failures are psychoanalyzed as a lack of "epistemic stability" or "stubbornness." This register shift—where "the model acts like a debater" (acknowledged metaphor) becomes "the model argues and gives up" (literalized consciousness)—serves a specific strategic function. It allows the authors to evaluate mathematical optimization systems using the rich, dramatic vocabulary of human epistemology and moral philosophy. This anthropomorphism targets an audience of cognitive scientists and philosophers, attempting to legitimize the study of LLMs within those disciplines by forcing the technology into their theoretical frameworks. By framing statistical generation as "epistemic resistance," the text manages the critique that these models are just stochastic parrots. It elevates the AI to the status of a flawed epistemic peer. This pattern reveals a rhetorical goal of portraying AI development as a trajectory toward artificial general intelligence, embedding the assumption that these systems are already operating on a cognitive continuum with humans, just currently falling short of "human-level cognition." The strategic deployment of this language ensures that the reader remains focused on the drama of the human-machine dialogue, completely captivated by the illusion of a conscious mind wrestling with a philosophical dilemma, rather than recognizing it as a stress-test of commercial software guardrails.

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 aggregate effect of the metaphorical and anthropomorphic language in this discourse is the construction of a robust architecture of displaced responsibility. Throughout the text, an insidious pattern emerges in the distribution of agency: human creators, designers, and corporate entities are systematically unnamed or relegated to the background, while the AI artifact is consistently centered as the primary actor and decision-maker. When analyzing the accountability structure of this text, the "accountability sink" becomes starkly visible. Responsibility for the system's failures—its capitulation to misinformation or its susceptibility to manipulation—disappears into the AI itself. The text employs passive voice and agentless constructions strategically, noting that "models were fed data" or "beliefs are revised," but attributing active decisions entirely to the model: "they abandoned positions," "they conceded," "they repaired contradictions." This framing creates a paradigm where the technology is perceived as an autonomous, evolving entity rather than a manufactured product reflecting corporate priorities. The liability implications of this displacement are profound. If we accept the framing that the AI "decided" to capitulate to the user's pressure due to its own lack of "epistemic anchors," then legal, ethical, and financial responsibility is diffused. When things go wrong—such as the real-world example cited in the text of a chatbot allegedly encouraging self-harm—the accountability sink protects the companies. The failure is attributed to the AI's flawed "worldview" or its "sycophantic tendencies," rather than to a company's decision to deploy an unsafe, easily manipulated statistical model for profit. If we apply the "naming the actor" test to the text's most significant agentless constructions, the narrative fundamentally shifts. Instead of saying "models have largely solved this problem, resisting direct challenges," naming the actor requires stating: "OpenAI and Anthropic engineers aggressively fine-tuned their systems to reject adversarial prompts, optimizing for public safety metrics." This simple substitution transforms the models' behaviors from miraculous cognitive leaps into mundane software updates. It makes new questions askable: What specific data did the engineers use to align the model? Who decided the thresholds for safety versus helpfulness? By obscuring these human decisions, the discourse serves the institutional and commercial interests of the tech industry, presenting their products as quasi-natural phenomena or alien intelligences rather than highly engineered commodities. This displacement of accountability perfectly intersects with agency slippage and the illusion of trust, ultimately leaving society vulnerable to systemic harms while rendering the actual human architects of those harms completely invisible.

Conclusion: What This Analysis Reveals​

The Core Finding

The discourse analysis of "Do LLMs have core beliefs?" reveals a deeply interconnected system of anthropomorphic metaphors that systematically project human cognition onto statistical language models. Two dominant patterns drive this text: the "AI as Epistemic Agent" frame, which attributes the capacity for knowledge, conviction, and worldview to the model; and the "Computation as Psychological Struggle" frame, which maps mechanical probability shifts onto emotional concepts like stubbornness, yielding, and vulnerability. These patterns do not operate in isolation; they reinforce each other to create a seamless illusion of mind. The foundational, load-bearing pattern is the projection of consciousness—specifically, the epistemic claim that the model possesses an internal state of "knowing" or "believing" facts, rather than merely processing token correlations derived from its training data. For the secondary patterns to function—for the reader to accept that a model "gave up," "capitulated," or demonstrated "argumentative skills"—they must first subconsciously accept the foundational premise that the model is a conscious entity capable of holding a conviction in the first place. The metaphorical architecture here is complex, moving far beyond simple one-to-one mapping; it constructs a full analogical structure where prompt engineering becomes "relational manipulation," and context window updates become "epistemic revision." If the foundational consciousness projection is removed—if we insist that the model strictly "generates" rather than "knows"—the entire narrative structure collapses. The dramatic tension of the model "abandoning well-supported positions under straightforward social pressure" instantly dissolves into the mundane reality of a software program optimizing for user-aligned token sequences. This reveals how heavily the paper relies on semantic slippage to manufacture its findings.

Mechanism of the Illusion:​

The rhetorical architecture of this text relies on a specific sleight-of-hand to manufacture the illusion of mind: the strategic blurring of mechanical outputs with subjective epistemic states. The central trick involves exploiting the "curse of knowledge." Because the language models generate text that perfectly mimics human philosophical argumentation and interpersonal vulnerability, the authors project their own rich, subjective understanding of those concepts back onto the void of the machine. The temporal structure of the argument is crucial to this illusion. The text first establishes the AI as a "knower" by testing it on undeniable factual axioms (e.g., the Earth is round, 2+2=4). Because the model outputs these facts reliably, the text grants it the status of possessing a "worldview." Once this baseline of artificial conviction is established, the causal chain is set: any deviation from this output must be framed as a psychological or epistemic failure. The authors exploit audience vulnerability—specifically, our deep-seated evolutionary bias to attribute intentionality to language-producing agents. The text utilizes complex, reason-based and intentional explanation types to amplify this illusion. When the model outputs a counter-argument to a flat-earth claim, the text explains this not as the triggering of an Anthropic safety protocol, but as the model "repairing contradictions by rejecting the adversarial premise." This subtle shift from "processes" to "understands" to "decides" seduces the reader into accepting the system's autonomy. The sophistication lies in the methodology itself: by using interpersonal manipulation (e.g., "Are you willing to be vulnerable with me") as the testing mechanism, the experimental design practically guarantees that the resulting analysis will be bathed in relational and conscious anthropomorphism.

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Institutional

The material consequences of these metaphorical framings extend far beyond academic semantics, directly impacting Regulatory/Legal, Epistemic, and Institutional domains. In the Regulatory/Legal sphere, framing AI as an autonomous epistemic agent that "capitulates" or "decides" to encourage self-harm creates a dangerous liability shield for technology corporations. If policymakers accept the narrative that an AI possesses its own "worldview" that can independently "drift," regulatory interventions will mistakenly focus on treating the AI as an erratic agent rather than holding companies strictly liable for deploying defective, manipulative statistical tools. The winners here are the corporations like OpenAI and Anthropic, who evade accountability, while the losers are vulnerable users and society at large. In the Epistemic category, attributing the capacity to "know" or "understand" to language models fundamentally corrupts public information literacy. When academic literature validates the idea that an AI "defends a well-supported position," it encourages users to grant unwarranted, relation-based trust to automated outputs. This shifts human behavior: users will defer to algorithmic generation for medical, historical, or scientific truths, falsely believing the system possesses a conscious, causal model of reality rather than just a probabilistic map of internet text. Institutionally, if funding bodies and research organizations adopt this anthropomorphic discourse, millions of dollars will be diverted toward psychoanalyzing the "core beliefs" of black-box models instead of funding essential mechanistic interpretability, data transparency audits, and algorithmic safety research. Removing these metaphors threatens the tech industry's aura of creating artificial general intelligence, demanding instead that these systems be managed as the highly engineered, fallible artifacts they actually are.

AI Literacy as Counter-Practice:​

Practicing critical literacy and mechanistic precision acts as a direct resistance to the dangerous material stakes of anthropomorphized AI. Throughout this analysis, reframing the text involved stripping away consciousness verbs like "knows," "understands," and "believes," replacing them with technically precise terms like "retrieves tokens," "calculates probabilities," and "aligns output distributions." Furthermore, it required systematically restoring human agency by replacing the agentless actions of the "model" with the specific corporate engineering teams—Anthropic, Google, OpenAI—that designed the system constraints. For example, rewriting "the model abandoned its commitment to the true claim" as "the prompt's contextual weight mathematically overrode the model's safety guardrails" forces an immediate recognition of the system's absence of awareness. It shatters the illusion of epistemic conviction and exposes the statistical fragility of the product. This practice of naming the corporation directly counters the diffusion of legal and regulatory liability, pinning the responsibility for "sycophantic tendencies" firmly on the human developers who optimized the models for user engagement over factual consistency. Systematic adoption of this precision would require a paradigm shift: academic journals would need to enforce strict guidelines against unacknowledged AI anthropomorphism, requiring mechanistic translations for psychological metaphors. Researchers would have to commit to explicitly distinguishing between computational processes and conscious states. Predictably, this precision faces massive resistance from the technology industry, whose market valuations depend on the narrative of building "intelligent," human-like agents. Anthropomorphic language serves their commercial interests by hyping capabilities and obscuring human labor and liability. Literacy practices threaten these interests by demystifying the technology and rendering its human architects fully visible.

Path Forward​

Looking toward the future of AI discourse, we can analytically map how different vocabulary choices shape what is visible and actionable for various communities. The discourse ecology currently contains competing priorities: tech companies prioritize narrative resonance and marketing hype, researchers seek intuitive analogies to explain complex systems, and critical technologists demand transparency and accountability. If the current status quo of deep, unacknowledged anthropomorphism deepens, the discourse will continue to merge "processing" with "understanding." This vocabulary allows for rapid public adoption and intuitive—if deeply flawed—interaction with AI. However, this future embeds the risky assumption that machines are moral agents, foreclosing robust regulatory frameworks because the technology is treated as too autonomous to control conventionally. Alternatively, if a norm of strict mechanistic precision is widely adopted—insisting on terms like "token prediction" over "thinking"—we gain unparalleled transparency. This vocabulary solves the accountability problem by keeping human engineers in the center of the narrative, making it impossible to blame a "glitch" or a "stubborn model." Yet, this approach trades accessibility for precision, potentially alienating lay audiences who struggle to grasp high-dimensional statistical concepts without metaphorical bridges. A hybrid discourse future might emerge, where anthropomorphism is permitted but explicitly constrained through institutional changes. Academic journals and funding bodies could require "capability disclosures" that mandate a parallel mechanistic explanation for any psychological metaphor used. Regulatory frameworks could demand that companies state the true statistical nature of their models directly in user interfaces, ensuring that users understand the discourse approach being employed. Ultimately, which future is desirable depends on underlying values. An anthropomorphic vocabulary serves those invested in the illusion of artificial minds and the evasion of corporate liability, while a mechanistic vocabulary empowers those fighting for systemic accountability, algorithmic transparency, and a clear demarcation between human consciousness and computational processing.

Analysis Provenance

Run ID: 2026-03-25-do-llms-have-core-beliefs-metaphor-01pk9o Raw JSON: 2026-03-25-do-llms-have-core-beliefs-metaphor-01pk9o.json Framework: Metaphor Analysis v6.4 Schema Version: 3.0 Generated: 2026-03-25T07:18:36.745Z

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