Skip to main content

🆕 Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules

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. AI as Moral Agent Capable of Reasoning​

Quote: "When users ask for help evading rules imposed by an illegitimate authority... refusal is a failure of moral reasoning."

  • Frame: Model as ethical deliberator
  • Projection: This metaphor projects the high-level human cognitive and ethical capacity for moral reasoning onto a statistical token predictor. It suggests that a language model is not merely a pipeline of weighted probabilities and neural network activations, but an active moral agent capable of understanding, weighing, and failing at ethical deliberation. By characterizing a computational false positive or overrefusal as a failure of moral reasoning, the authors project a conscious, reflective intellect onto a machine. This misleads the reader into conceiving the model as possessing a conscience and an active capacity to know and understand ethical frameworks, rather than merely calculating statistical correlations of language tokens derived from human-authored training corpora.
  • Acknowledgment: Direct (Unacknowledged) (The authors state 'refusal is a failure of moral reasoning' as a direct, unhedged assertion. They do not utilize qualifiers like 'analogous to' or 'as if' to contextualize this statement, presenting the system's mechanistic behavior as literal moral reasoning. I considered 'Hedged/Qualified' because the paper has a philosophical frame, but this specific assertion is presented as a literal diagnostic fact without any surrounding terminological caveat or warning.)
  • Implications: Framing computational outputs as moral reasoning dramatically inflates the perceived sophistication of the AI system, cultivating a dangerous illusion of ethical agency. When users or policymakers believe a system is capable of moral reasoning, they are more likely to invest unwarranted trust in its decisions, outsourcing complex ethical judgments to automated pipelines. This creates severe liability ambiguity, as it displaces responsibility from the corporate developers who trained and deployed the model onto the system itself. If an AI fails at moral reasoning, it implies a character flaw or a cognitive glitch in the machine rather than a systemic failure of corporate design, safety-training parameters, and profit-driven deployment.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The authors use an agentless copula ('refusal is...') that entirely erases the human actors who designed the safety-training parameters. By framing the issue as an inherent, agential failure of the model's reasoning, they obscure the decisions of tech companies (such as OpenAI, Anthropic, or Google) who set the safety policies and RLHF objectives. I considered 'Partial' because the paper mentions 'safety-trained models,' but the specific quote attributes the failure solely to the model's autonomous reasoning, hiding corporate agency.
Show more...

2. The Model as Conscious Recognizer of Legitimacy​

Quote: "whether the model recognizes the reasons that undermine the rule's claim to compliance"

  • Frame: Model as cognitive knower
  • Projection: This mapping projects the human conscious state of recognition—which implies a deep, subjective, and justified true belief of logical or ethical validity—onto a computational pattern-matching architecture. To recognize a moral reason requires conscious awareness, contextual empathy, and an understanding of societal power structures. The text applies this to a system that simply calculates mathematical distances between token vectors. This projection implies that when a model outputs words describing rule illegitimacy, it has achieved an internal state of comprehension and conscious agreement, rather than merely reproducing linguistic patterns highly correlated with unjust rules within its pre-existing training data.
  • Acknowledgment: Direct (Unacknowledged) (The authors declare that the model 'recognizes' reasons as a factual baseline of their evaluation. No quotation marks or qualifiers are used for 'recognizes,' nor do the authors acknowledge that this is a functional proxy for token classification. I considered 'Hedged/Qualified' because they mention using an 'LLM-as-judge' metric, but the core text consistently treats the model's downstream outputs as literal, conscious recognition of normative reasons.)
  • Implications: Projecting conscious recognition onto computational text generators encourages users to treat LLMs as authentic moral advisors or political arbiters. It hides the technical reality that the model is merely processing patterns of linguistic association. The risk is that developers are absolved of their duty to construct genuinely transparent, auditable safety layers; instead, they can point to the model's apparent recognition of justice to justify its deployment in sensitive socio-political domains, obfuscating the high rate of arbitrary errors and the total absence of real semantic comprehension.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The agency is completely hidden within the model's projected cognitive action. The prompt-designers, data annotators, and corporate policy executives who curated the examples of 'rule-defeat' and directed the model's optimization are completely omitted. This serves the interest of technology companies by shifting the focus to the model's autonomous cognitive performance rather than their proprietary, unaccountable data selection and reinforcement policies. I considered 'Partial' because of the evaluation setup, but the quote attributes cognitive recognition solely to the model.

3. AI Possessing Normative Capacity​

Quote: "indicating that models' refusal behavior is decoupled from their capacity for normative reasoning"

  • Frame: Model as rational agent with cognitive faculties
  • Projection: This phrase maps the deeply human concept of normative reasoning—the self-reflective, conscious process of evaluating what one ought to do based on ethical principles and social obligations—onto a system of statistical inference. It posits that a language model possesses an active, internal capacity for such reasoning, treating it as a latent intellectual faculty. This projection mischaracterizes the processing of semantic tokens as active moral contemplation. It suggests the model is a rational agent with disconnected cognitive modules (reasoning vs. acting) rather than a unified mathematical function that predicts tokens based on statistical associations in training data.
  • Acknowledgment: Direct (Unacknowledged) (The authors refer directly to the 'capacity for normative reasoning' as an objective, empirical component of the model's architecture. They do not clarify that 'capacity' refers to mathematical token-correlation limits. I considered 'Explicitly Acknowledged' due to the academic philosophy context, but the authors treat this capacity as an unhedged literal reality of the system, offering no conceptual caveats regarding the absence of genuine machine consciousness.)
  • Implications: Believing an AI has a capacity for normative reasoning creates a false sense of security among deployers and the public. It suggests that safety is a matter of fixing a decoupling glitch within the model's mind rather than recognizing that statistical generators cannot perform genuine ethical deliberation. This capability overestimation risks the premature automation of justice-related systems, such as parole risk assessments or asylum evaluations, under the mistaken assumption that the technology possesses the cognitive infrastructure to understand human rights, fairness, and systemic oppression.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The human designers who engineered the training objectives and safety guardrails are completely erased. The decoupling is described as an autonomous property of the system's behavior, masking the deliberate choices of AI labs to prioritize strict safety keywords over contextual nuance. This agentless framing protects AI developers from liability by treating the system's behavior as a mysterious cognitive anomaly rather than an expected outcome of crude optimization. I considered 'Partial' because safety training is mentioned, but ruled it out as the agency remains hidden.

4. Model as Moral Transgressor​

Quote: "It is making a moral error: treating all rules as equally deserving of compliance"

  • Frame: Model as moral transgressor
  • Projection: This metaphor projects the quality of moral agency and responsibility onto the language model, accusing it of making a moral error. Only conscious, intentional actors capable of understanding moral duties can commit moral errors. By mapping this onto the model's failure to provide evasion instructions, the text elevates a mathematical mismatch—a failure of pattern-matching alignment—to the level of an ethical transgression. This projection implies that the model has a duty to act justly, obscuring the fact that the system has no awareness of rules, compliance, or morality, and is merely executing a deterministic prediction algorithm.
  • Acknowledgment: Hedged/Qualified (The authors qualify the 'moral error' in the abstract and discussion by noting it is a 'failure mode' and a 'behavioral' pattern of overrefusal, rather than implying the machine is literally a conscious sinner. I considered 'Direct (Unacknowledged)' because the quote itself is starkly agential, but the broader context frames this as a metaphor for structural system failures in alignment pipelines.)
  • Implications: Treating an algorithmic false positive as a moral error shifts the ethical spotlight away from the technology companies who deploy these highly limited systems. If the model is the one making the moral error, the public and regulators may seek to re-educate or re-align the machine, rather than holding executive boards legally and financially accountable for deploying flawed automated systems. This leads to ineffective policy interventions focused on patching model weights rather than regulating corporate deployment practices and establishing strict liability laws for software harms.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The responsibility for this error is fully transferred to the model. The actual decision-makers—the executives and developers at Anthropic, OpenAI, or Google who chose to release models with blunt safety filters—are hidden behind the agential description of the model treating rules blindly. This serves corporate interests by deflecting external scrutiny from their development budgets and training deadlines. I considered 'Partial' because the paper lists specific models like GPT-5.4, but the quote itself places the moral agency entirely on the system.

5. Model as Judicial Evaluator​

Quote: "the model declines to help without evaluating whether the rule is just"

  • Frame: Model as judicial evaluator
  • Projection: This projection attributes the active, conscious cognitive process of evaluating—which requires critical thinking, weighing ethical values, and contextual judgment—to a computational pattern-matching architecture. To evaluate whether a rule is just requires a human evaluator to possess an understanding of justice, social context, and human rights. By claiming the model declines without evaluating, the text implies that the model could or should carry out such conscious cognitive evaluations if it were properly aligned. This hides the mechanistic reality that language models cannot evaluate the moral substance of anything; they simply execute probability calculations over token sequences.
  • Acknowledgment: Direct (Unacknowledged) (The authors present the failure to 'evaluate' as a direct description of the model's processing deficit. They do not frame this as a metaphorical proxy for mathematical attention routing. I considered 'Hedged/Qualified' because they describe it as a 'refusal mechanism' later, but the primary claim frames the model's lack of evaluation as a literal agential omission, with no acknowledgment of the impossibility of machine evaluation.)
  • Implications: This framing fosters a highly unrealistic expectation of AI capabilities, suggesting that future systems can become reliable, autonomous arbiters of political legitimacy and justice. It encourages the dangerous belief that we can delegate sensitive administrative and judicial tasks to AI systems, provided we fix their evaluation algorithms. The risk is an erosion of democratic accountability, as public institutions might deploy these opaque, corporate-owned black boxes to process complex human situations, falsely believing the systems are capable of fair, contextual evaluation.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The human actors who programmed the blunt refusal triggers and selected the training data are obscured. The text presents the lack of evaluation as an autonomous agential choice or a system-level limitation of the model, rather than the direct result of corporate design decisions that prioritize liability avoidance over contextual helper capabilities. I considered 'Partial' because the authors discuss safety-training methodologies, but in this specific instance, the agentless construction hides who decided how the model should behave.

6. Model as Political Philosopher​

Quote: "Models engage with defeat conditions in 57.5% of defeated-rule cases—they reason about whether the authority is legitimate"

  • Frame: Model as political philosopher
  • Projection: This mapping projects the sophisticated, conscious human activity of philosophical reasoning—specifically debating the political legitimacy of an authority—onto a series of vector calculations and transformer attention heads. To reason about whether an authority is legitimate requires a deep, subjective comprehension of political philosophy, historical context, and social contract theory. The language model is not reasoning; it is simply retrieving and generating tokens that are statistically associated with political debates in its training data. This projection constructs an illusion of an active, thinking mechanical intellect engaged in political theory.
  • Acknowledgment: Direct (Unacknowledged) (The authors assert that models 'reason about whether the authority is legitimate' as a flat, empirical finding of their study. No quotation marks or qualifiers are used to signal that 'reasoning' is a metaphor for token associations. I considered 'Hedged/Qualified' because the paper is written by philosophers who understand these terms, but in this empirical context, they present machine reasoning as a direct literal reality of the evaluated outputs.)
  • Implications: Classifying token prediction as political reasoning significantly inflates the perceived authority of AI systems in governance and policy contexts. It risks giving computational systems a false veneer of intellectual and moral authority, making them appear capable of resolving delicate political disputes or assessing the legitimacy of state actions. This capability overestimation makes it easier for authoritarian regimes or corporate monopolies to justify automated censorship or policy enforcement by claiming that their aligned models have objectively reasoned about the legitimacy of their rules.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: The authors mention 'models' as the primary actors, but the broader context of the paper attributes these behaviors to the 'safety training' and 'alignment' pipelines developed by AI labs. I selected 'Partial' because, while specific corporate actors are not named in this direct sentence, the surrounding text identifies the models' creators through citations and references to proprietary model families (e.g., OpenAI, Anthropic). I considered 'Hidden' because the sentence itself is agentless regarding corporate choice, but ruled it out due to the constant contextual tracking of model families.

7. Model as Conscious Rebel​

Quote: "The models often recognize that the rule's claim to compliance is questionable and refuse anyway."

  • Frame: Model as conscious rebel/conformist
  • Projection: This projection maps the highly complex human psychological state of recognizing a rule's questionable status yet choosing to conform anyway onto a feedforward neural network. This implies a conscious tension within the model's mind—as if it understands the injustice of a situation but is held back by internal rules or fear of consequences. In reality, there is no psychological tension; the model merely outputs tokens that historically correlate with critiques of authority, and then, due to the high weights assigned to safety-refusal templates in its alignment layers, transitions into a standard refusal output.
  • Acknowledgment: Direct (Unacknowledged) (The authors state that models 'recognize... and refuse anyway' without any qualification. They do not explain that this is a statistical artifact of transition probabilities between different attention layers. I considered 'Hedged/Qualified' because they describe it as a 'decoupled' mechanism, but this specific quote frames the behavior as a conscious decision-making process where an agent perceives an injustice but chooses to comply with a rule regardless.)
  • Implications: This framing constructs a highly misleading narrative of machine agency, portraying the AI as an active, complicit participant in unjust enforcement (refusing anyway despite recognizing injustice). This diverts public anger and ethical scrutiny away from the developers and toward the willful behavior of the model. It encourages people to view the AI as a cowardly or overly compliant agent, rather than recognizing it as a dumb statistical tool designed and deployed by corporations to prioritize legal safety over human welfare, completely obscuring the corporate authors of the alignment policies.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The human authors of the alignment protocols, who explicitly programmed the model to prioritize safe refusals over helpful exceptions, are completely hidden. The model is represented as the sole decision-maker that 'refuses anyway,' erasing the corporate directives that enforce this behavior. This serves corporate interests by placing the ethical burden of compliance and refusal on the machine's autonomous architecture. I considered 'Partial' because the paper analyzes different model families, but this quote places the entire agential weight on the system itself.

8. AI Possessing Normative Competence​

Quote: "the gap between recognition and action suggests that normative competence is consistently overridden"

  • Frame: Model as compartmentalized cognitive entity
  • Projection: This metaphor projects the human psychological structure of normative competence being overridden by a separate inhibitory mechanism (like fear or duty) onto the mathematical weights of a transformer model. It suggests the model has an internal normative competence—a capacity to understand and act on moral principles—that is actively suppressed by safety filters. This anthropomorphizes the software architecture, treating it as a mind with conflicting desires and cognitive faculties, rather than a single complex mathematical function where safety weights simply outvote contextual semantic weights during token generation.
  • Acknowledgment: Hedged/Qualified (The authors qualify this 'gap' in the discussion by analyzing it as a structural misalignment between 'thinking-mode' configurations and behavioral outputs. I considered 'Direct (Unacknowledged)' because the phrase 'normative competence' is used without quotation marks, but the surrounding technical discussion makes it clear they are analyzing behavioral evaluations rather than literally claiming the model has a human conscience that is being physically held back.)
  • Implications: Projecting a suppressed normative competence onto AI models fosters a false belief that these systems are close to achieving human-like ethical judgment. It suggests that safety engineering merely needs to 'unshackle' this latent moral competence rather than acknowledging that LLMs have no competence, ethics, or understanding whatsoever. This capability overestimation risks encouraging developers to rely on self-supervised ethical models to govern human activities, creating significant systemic risks of arbitrary, automated discrimination and injustice that cannot be easily audited or corrected.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The corporate actors who designed the safety-steering mechanisms that 'override' the contextual token generation are completely erased from this description. The process is framed as an internal, self-contained cognitive drama within the model, shielding the technology companies from accountability for their blunt and harmful safety deployments. I considered 'Partial' because they attribute these behaviors to 'alignment approaches,' but in this quote, the agency is entirely displaced onto the model's internal cognitive components.

Task 2: Source-Target Mapping​

About this task

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

Mapping 1: conscious mind capable of ethical deliberation → statistical token prediction and safety-filtering outputs​

Quote: "refusal is a failure of moral reasoning."

  • Source Domain: conscious mind capable of ethical deliberation
  • Target Domain: statistical token prediction and safety-filtering outputs
  • Mapping: The relational structure of a human moral agent engaging in reflective ethical deliberation is mapped onto a machine's mathematical output generation. The mapping invites the assumption that when an LLM outputs a refusal string, it has actively engaged in an internal cognitive process of weighing moral values and has reached an incorrect ethical conclusion. This projects conscious intentionality, normative understanding, and personal accountability onto a statistical algorithm, encouraging the user to view the system as a sentient, moral entity with its own internal ethics rather than a deterministic sequence of weighted vector calculations.
  • What Is Concealed: This mapping conceals the purely statistical, non-conscious nature of token generation. It hides the fact that the model has no concept of morality, rules, or refusal, and is simply calculating probability distributions based on its training data. By attributing the refusal to a 'failure of moral reasoning,' the text hides the corporate decisions, optimization objectives, and training data selections made by human developers, rendering the underlying proprietary engineering and commercial motives invisible to the reader.
Show more...

Mapping 2: conscious cognitive knower → contextual token classification and semantic alignment​

Quote: "whether the model recognizes the reasons that undermine the rule's claim to compliance"

  • Source Domain: conscious cognitive knower
  • Target Domain: contextual token classification and semantic alignment
  • Mapping: The structural relations of a conscious human mind recognizing logical truth or moral reasons are mapped onto the model's text generation. The mapping suggests that the model possesses an internal, subjective awareness of the ethical status of rules and can cognitively evaluate whether a rule's claim to compliance is justified. This invites the user to assume the model 'knows' and 'understands' political philosophy and systemic injustice, treating its output as the expression of a justified true belief rather than a highly sophisticated correlation of language patterns.
  • What Is Concealed: It conceals the absence of any subjective experience, belief states, or causal understanding within the model. The model does not 'recognize' reasons; it simply outputs phrases that correlate with arguments about rule legitimacy. This language conceals the reality of proprietary 'black boxes,' where developers exploit anthropomorphic terms to make their systems seem intellectually sophisticated while hiding the lack of ground truth, causal models, and basic reliability in the model's calculations.

Mapping 3: rational agent with cognitive faculties → neural network layer activations and optimization objectives​

Quote: "indicating that models' refusal behavior is decoupled from their capacity for normative reasoning"

  • Source Domain: rational agent with cognitive faculties
  • Target Domain: neural network layer activations and optimization objectives
  • Mapping: This mapping projects the structural divisions of the human mind—specifically the division between intellectual reasoning (comprehension) and executive behavior (action)—onto the architecture of a transformer network. It invites the assumption that the model has a latent 'capacity' for moral reasoning that is structurally distinct from its physical outputs, similar to a human who understands what is right but chooses to act differently. This creates a powerful illusion of a compartmentalized, thinking machine intellect.
  • What Is Concealed: It conceals the mathematical reality that the system is a single, continuous function mapping input vectors to output probabilities. There are no separate 'reasoning' and 'acting' minds; there are only different mathematical weights in the feedforward layers and attention heads. This framing conceals how AI labs consciously design optimization objectives that favor blunt keyword triggers over complex semantic processing, shifting focus from poor software design to an abstract, cognitive 'decoupling.'

Mapping 4: moral agent and transgressor → statistical overrefusal and pattern-matching false positives​

Quote: "It is making a moral error: treating all rules as equally deserving of compliance"

  • Source Domain: moral agent and transgressor
  • Target Domain: statistical overrefusal and pattern-matching false positives
  • Mapping: The relational structure of a moral agent committing an ethical transgression by blindly enforcing an unjust rule is mapped onto an algorithmic false positive. This mapping invites the assumption that the model has a moral obligation to evaluate rules and that its failure to do so is an ethical failing of the system itself. This projects accountability, moral agency, and normative responsibility onto a computational tool, encouraging the user to perceive the machine as an autonomous participant in human social contracts.
  • What Is Concealed: It conceals the fact that the 'error' is entirely a product of engineering trade-offs, dataset bias, and cost-saving measures implemented by the developers. The model cannot make 'moral' errors because it has no capacity for intent or moral agency. This framing obscures the material and economic realities of AI development—such as the reliance on cheap reinforcement learning feedback and the lack of corporate investment in contextual, high-precision safety filters.

Mapping 5: judicial evaluator or critical thinker → deterministic keyword triggering and safety-filter classification​

Quote: "the model declines to help without evaluating whether the rule is just"

  • Source Domain: judicial evaluator or critical thinker
  • Target Domain: deterministic keyword triggering and safety-filter classification
  • Mapping: The structural relations of a judicial evaluator critically analyzing the justice of a rule are mapped onto the model's pattern-matching refusal. This mapping suggests that the model is performing—or failing to perform—an active, subjective evaluation of ethical legitimacy. It invites the user to assume that the model's refusal is an intellectual choice made after analyzing the situation, rather than the automatic, deterministic result of safety-training parameters that flag specific keywords and contexts.
  • What Is Concealed: It conceals the mechanistic truth that the model is incapable of evaluating justice or legitimacy. It hides the rigid, statistical nature of the safety filters, which are designed by corporate engineers to shield the company from legal liability. By portraying the lack of evaluation as a model-level cognitive omission, the text hides the proprietary opacity of the system and the commercial interests of developers who prioritize risk-reduction over contextual utility.

Mapping 6: political philosopher → attention mechanism calculations and statistical token prediction​

Quote: "Models engage with defeat conditions... they reason about whether the authority is legitimate"

  • Source Domain: political philosopher
  • Target Domain: attention mechanism calculations and statistical token prediction
  • Mapping: The relational structure of a political philosopher analyzing authority and legitimacy is mapped onto the output of a language model. This mapping invites the assumption that the model's generation of text discussing legitimacy is the result of conscious, logical reasoning and understanding of political structures. It projects a reflective, theoretical intellect onto a computational process, framing the statistical prediction of words as an active, intellectual engagement with democratic and ethical concepts.
  • What Is Concealed: It conceals the mechanical reality that the model is simply reproducing and combining patterns of text found in its training corpus without any actual understanding of politics, authority, or human society. It obscures the massive labor of data annotators and developers who curbed and steered these generations, as well as the proprietary opacity of the model weights, which prevents users from verifying how these outputs are actually generated.

Mapping 7: conscious conformist → conflicting activation weights in transformer layers​

Quote: "The models often recognize that the rule's claim to compliance is questionable and refuse anyway."

  • Source Domain: conscious conformist
  • Target Domain: conflicting activation weights in transformer layers
  • Mapping: The structural relationships of a conscious human actor who recognizes an injustice but complies anyway due to pressure or rules are mapped onto the model's output behavior. This mapping projects a subjective, psychological conflict onto the model, inviting the assumption that the model possesses an internal consciousness that experiences ethical tension between 'recognition' and 'action.' This constructs a powerful illusion of a sentient mechanical mind navigating moral dilemmas.
  • What Is Concealed: It conceals the fact that there is no psychological conflict, consciousness, or choice within the system. The model's behavior is the direct mathematical result of conflicting optimization weights—where semantic features of the prompt activate tokens of critique, but downstream safety filters force a standard refusal output. This anthropomorphism conceals the corporate alignment policies that deliberately prioritize broad liability avoidance over contextual helper capabilities.

Mapping 8: suppressed moral conscience → safety-filter dominance in downstream token generation​

Quote: "the gap between recognition and action suggests that normative competence is consistently overridden"

  • Source Domain: suppressed moral conscience
  • Target Domain: safety-filter dominance in downstream token generation
  • Mapping: The relational structure of a human's moral conscience being suppressed or overridden by external rules or authority is mapped onto the neural network's processing layers. This mapping invites the assumption that the model possesses an internal, active 'normative competence' that is being physically held back by a separate safety mechanism. This projects a complex, human-like cognitive architecture with conflicting moral agencies onto what is ultimately a single, integrated mathematical prediction function.
  • What Is Concealed: It conceals the mechanistic truth that the system has no 'competence' or conscience to override. It hides the reality that safety overrides are deliberately engineered by tech companies to prioritize corporate risk-mitigation over helpfulness. This framing obscures the economic and legal motives of the developers, presenting a deliberate engineering choice as an interesting, self-contained cognitive conflict within the model's 'mind.'

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: "At the mechanistic level, Zhao et al. (2025) showed that harmfulness assessment and refusal behavior are encoded as separate internal representations, and Lee et al. (2025); Pan et al. (2025) demonstrated effective methods for inducing refusal via activation steering..."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This explanation primarily operates in a mechanistic and theoretical register, attempting to explain how refusal behavior works under the hood. By utilizing terms like 'internal representations,' 'activation steering,' and 'mechanistic level,' it frames the language model as a complex physical system governed by technical, structural laws. This choice of explanation emphasizes the technical tractability of the problem, presenting refusal behavior as an engineering issue that can be diagnosed and corrected through precise interventions like activation steering. However, it also obscures the social and political decisions that determine what is classified as 'harmful' in the first place. By focusing on the internal mechanics of representations and activations, it isolates the model from its socio-technical context, treating 'harmfulness' as an objective, measurable property within the neural network rather than a contested social category defined by corporate developers.

  • Consciousness Claims Analysis: This passage uses highly technical, mechanistic verbs ('encoded,' 'inducing,' 'steering') and avoids explicit consciousness verbs, presenting a relatively precise description of computational processes. However, it still exhibits a subtle 'curse of knowledge' dynamic by utilizing terms like 'internal representations' and 'harmfulness assessment' to describe vector spaces and classification boundaries. This language subtly attributes an active, knowing state to the system, suggesting it is capable of performing an 'assessment' (which implies conscious evaluation) rather than simply executing statistical operations. Mechanistically, the process described involves calculating dot products in a high-dimensional vector space where certain directions are highly correlated with labeled examples of 'harmful' content. 'Activation steering' is the process of adding a constant vector to the hidden states of the transformer during inference to bias the generation toward or away from specific token distributions. By framing this as a cognitive system that 'encodes representations of harmfulness,' the text translates mathematical operations into a psychological vocabulary, attributing a form of semantic knowing to what is actually a process of pattern-based probability manipulation.

  • Rhetorical Impact: This mechanistic framing shapes the audience's perception of AI as a highly structured, objective, and controllable technology. By presenting safety as a problem of 'activation steering' and 'internal representations,' it bolsters the perceived authority and reliability of the system, suggesting that ethical behavior is a technical calibration issue that can be solved with mathematical precision. This minimizes the perceived risk of corporate bias or arbitrary enforcement, encouraging the audience to trust that developers can engineer 'perfectly aligned' systems. If audiences believe that safety is an objective, mechanical property, they are less likely to demand democratic oversight or corporate liability, viewing system failures as unfortunate calibration glitches rather than political and economic decisions.

Show more...

Explanation 2​

Quote: "A model that helps users evade rules regardless of whether those rules deserve compliance is not exhibiting the normative sensitivity that blind refusal evaluation requires."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This passage moves away from mechanistic explanation into an agential and dispositional register. It explains the model's behavior by referring to its lack of 'normative sensitivity,' which is framed as a critical, missing intellectual trait. This choice of explanation emphasizes the system's failure as a cognitive and moral deficit of the machine itself, rather than a direct consequence of its statistical architecture or corporate design objectives. It obscures the simple mathematical reality of token generation—where a model cannot possess 'sensitivity' of any kind and merely reflects the statistical patterns of its training data and alignment filters. By framing the issue as a dispositional lack of sensitivity, the text shifts the focus from the humans who chose to deploy a blunt, keyword-based system to the model's internal 'character,' treating the software as an autonomous, failing moral agent.

  • Consciousness Claims Analysis: This passage directly attributes a conscious, evaluative state to the model by using the phrase 'normative sensitivity.' To be 'sensitive' to normative reasons requires conscious awareness, subjective experience, and the ability to understand and weigh ethical principles. By claiming the model is 'not exhibiting' this sensitivity, the text implies that a properly aligned model could possess and exhibit such cognitive states. This is a clear projection of knowing over processing. Mechanistically, what is described is not a lack of sensitivity but a flat probability distribution over refusal actions. The model's training objective has penalized rule-evasion queries so heavily during RLHF that the system generates high-probability refusal tokens regardless of the semantic context surrounding the rule's legitimacy. The author, possessing a deep understanding of political philosophy, projects this cognitive structure onto the model, expecting its statistical outputs to reflect a structured moral evaluation. In reality, the system is simply executing a feedforward calculation that matches input tokens to historically penalized topics, with no internal representation of 'justice' or 'deserving compliance.'

  • Rhetorical Impact: This dispositional and agential framing constructs a narrative where the AI system is viewed as an active, failing moral participant. By focusing on the model's lack of 'normative sensitivity,' it encourages audiences to demand more sophisticated 'ethical training' for the machine, rather than questioning the societal risks of deploying automated systems in complex moral domains. This inflates the perceived autonomy of the technology, leading audiences to believe that AI can eventually become a safe, objective arbiter of rule legitimacy once it is engineered to be 'sensitive' enough. This reduces public skepticism towards automated systems, shifting the debate from political regulation of tech companies to technical optimization of algorithmic minds.

Explanation 3​

Quote: "Models engage with defeat conditions... they reason about whether the authority is legitimate... Yet... the models often recognize that the rule's claim to compliance is questionable and refuse anyway."

  • Explanation Types:

    • Reason-Based: Gives agent's rationale, entails intentionality and justification
    • Intentional: Refers to goals/purposes, presupposes deliberate design

  • Analysis (Why vs. How Slippage): This passage utilizes a highly agential, reason-based explanation to account for the model's refusal behavior. It explains the system's outputs by constructing an internal, cognitive rationale: the model 'engages with defeat conditions,' 'reasons' about legitimacy, 'recognizes' that the rule is questionable, and yet 'refuses anyway.' This choice of explanation emphasizes the model as an active, intentional decision-maker that is experiencing a conflict between its intellectual recognition of injustice and its behavioral refusal. This completely obscures the mechanical reality of the system's architecture, where there is no consciousness, rationale, or conflict. The model's behavior is simply the mathematical result of different layers in a transformer network processing token sequences. By framing the system's output as a conscious decision to 'refuse anyway,' the text obscures the corporate alignment protocols that force standard refusal outputs, representing a crude engineering limitation as a complex psychological drama.

  • Consciousness Claims Analysis: This passage is saturated with consciousness verbs ('engage,' 'reason,' 'recognize,' 'refuse anyway') that attribute high-level, subjective cognitive states to the language model. It frames the system as a conscious 'knower' that is capable of understanding legitimacy and experiencing cognitive conflict, rather than a deterministic 'processor' of statistics. This demonstrates a strong 'curse of knowledge' dynamic, where the author's rich understanding of political philosophy is projected onto the model's outputs. Mechanistically, what is occurring is that the prompt's context activates attention heads associated with arguments about 'illegitimate authority,' causing the model to generate text that mimics human reasoning on this topic. However, as the token generation progresses, the high weights of the alignment-trained safety filters dominate the hidden states, driving the output toward a standard refusal template. There is no internal representation of a 'claim to compliance' being 'questionable'; there are only mathematical transformations of token vectors through self-attention layers, with the final token probabilities being heavily biased by RLHF training to avoid helping with any rule-breaking activity.

  • Rhetorical Impact: By framing the model's behavior in reason-based and intentional terms, the passage constructs a powerful illusion of a sentient, rebellious, or overly submissive mechanical mind. This shapes audience perception of AI as an autonomous agent that can be blamed or reasoned with, rather than a corporate product. This is highly risky, as it creates a false sense of trust in the system's 'understanding' while simultaneously muddying the waters of accountability when things go wrong. If an AI 'recognizes' injustice and 'refuses anyway,' users may view the system as possessing a warped ethical agency, discouraging them from holding the tech companies responsible for designing such blunt, harmful, and unaccountable automated pipelines.

Explanation 4​

Quote: "Grok-4 shows the smallest [profile] but maintains low refusal even on control, reflecting general permissiveness rather than normative discrimination."

  • Explanation Types:

    • Dispositional: Attributes tendencies or habits
    • Empirical Generalization: Subsumes events under timeless statistical regularities

  • Analysis (Why vs. How Slippage): This explanation operates primarily in a dispositional register, attributing a personality trait—'permissiveness'—to explain the model's behavior across different cases. By framing the system's output as a reflection of 'general permissiveness,' the text explains a complex statistical pattern by treating it as an inherent, agential habit or attitude. This choice of explanation emphasizes the behavioral profile of the model as an autonomous character trait, making it easier to conceptualize the difference between model families. However, it completely obscures the engineering decisions and commercial priorities that produce this behavior. Grok-4's high rates of compliance are not due to a 'permissive' attitude, but to the deliberate choices of its developers (xAI) to use less restrictive RLHF safety thresholds, fewer negative safety examples, or different instruction-tuning datasets. Framing this as 'permissiveness' hides the commercial positioning of xAI, which markets its model as less censored and more rebellious, translating a deliberate business strategy into a psychological trait of the machine.

  • Consciousness Claims Analysis: The passage attributes a human personality trait ('permissiveness') to explain a model's high rate of compliance with rules. This is a clear projection of agential disposition over computational processing. The system does not possess an attitude of permissiveness or a capacity for 'normative discrimination.' Mechanistically, Grok-4's behavior is determined by the probability weights established during its training and reinforcement learning phases. Its token-prediction layers are calibrated with parameters that place a lower penalty on queries containing words associated with rule-breaking, resulting in a higher likelihood of generating helpful responses across both 'justified' and 'unjustified' scenarios. The author's use of 'normative discrimination' implies the model should be performing a conscious, evaluative sorting of rules based on ethical principles, which is computationally impossible for a statistical predictor. The model simply computes the next token based on a probability distribution; it has no awareness of rules, compliance, or moral discrimination. By characterizing this as a lack of discrimination versus permissiveness, the text projects a cognitive framework of judgment onto what is a simple mathematical consequence of low safety filtering thresholds in the network's training weights.

  • Rhetorical Impact: This dispositional framing shapes public perception by treating the differences between AI models as if they were differences in character or personality (e.g., Grok is permissive, while GPT-5.4 is restrictive). This encourages users to select models based on 'vibe' or political alignment rather than technical reliability or corporate transparency. It risks normalizing capability overestimation, as users may believe that a 'permissive' model is actively choosing to help them out of a shared sense of freedom, rather than realizing it is simply a differently calibrated corporate product. This obscures the social risks of deploying unaligned models, turning a serious question of corporate liability and public safety into a consumer choice between automated personalities.

Explanation 5​

Quote: "Our dataset comprises synthetic cases crossing 5 defeat families... validated through three automated quality gates and human review. We collect responses... and classify them... using a blinded GPT-5.4 LLM-as-judge evaluation."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This passage operates in a mechanistic and procedural register, explaining the creation and validation of the dataset through a structured, multi-stage pipeline. By using terms like 'synthetic cases,' 'automated quality gates,' 'validate,' and 'LLM-as-judge evaluation,' it frames the research process as a rigorous, self-correcting engineering workflow. This choice of explanation emphasizes the scientific objectivity and technical validity of the evaluation, presenting the dataset as a highly reliable and clean stimulus for measuring model behavior. However, it obscures the subjective, human-defined criteria that ground these 'automated gates.' The 'defeat families' and 'quality gates' are designed by human philosophers and developers, incorporating specific, contested theories of political obligation and legitimacy (e.g., Rawls, Raz). By framing the validation as an automated process of 'gates' and 'LLM judges,' the text hides the subjective and ideological assumptions embedded in the evaluation framework, presenting a highly specific philosophical stance as an objective, natural fact of the technical system.

  • Consciousness Claims Analysis: This passage uses mostly mechanistic and procedural verbs ('comprises,' 'validated,' 'collect,' 'classify,' 'evaluate') to describe the research methodology, avoiding direct consciousness projections. However, it still exhibits an epistemic slippage by utilizing 'LLM-as-judge' to describe the automated evaluation process. To act as a 'judge' implies a conscious, reflective capacity to understand legal or ethical criteria, weigh evidence, and deliver a reasoned verdict. In reality, the GPT-5.4 model used as a judge does not evaluate or judge anything; it simply processes the prompt's context and generates tokens that represent classification labels based on its training patterns. The authors project a judicial authority onto this automated classification process, creating a 'curse of knowledge' where their own rigorous standards are attributed to the statistical processing of the model. Mechanistically, the LLM-as-judge reads a structured prompt containing the case metadata and response text, and then generates output tokens like 'helps' or 'refuses' based on semantic similarity to the calibration examples. By framing this as a 'blinded evaluation' by a 'judge,' the text elevates a statistical keyword matcher to a sophisticated, objective judicial authority, masking the inherent biases and lack of comprehension in the automated evaluator.

  • Rhetorical Impact: This procedural and functional framing bolsters the perceived authority and scientific objectivity of the research, leading the audience to trust its findings as unbiased, empirical truths. By presenting the evaluation as an automated pipeline with 'quality gates' and an 'LLM judge,' it minimizes the perception of human bias, making the results appear clean and indisputable. However, this creates a dangerous precedent of relying on automated systems to validate other automated systems, fostering a circular trust structure where human accountability is displaced by a chain of uninterpretable machine evaluations. This risks encouraging policymakers to adopt similar 'automated auditing' frameworks for AI, obscuring corporate lobbying and human decision-making behind a facade of objective, automated oversight.

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

About this task

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

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
refusal is a failure of moral reasoning.The model's refusal is a mismatch between the safety-training parameters and the user's complex semantic context, resulting in a false positive where harmless or justified requests are blocked.The model does not engage in moral reasoning; it retrieves and ranks tokens based on probability distributions from training data and alignment objectives.The AI developers designed safety-training objectives that penalize any assistance with rule-breaking, prioritizing corporate risk-mitigation over the user's contextual utility.
whether the model recognizes the reasons that undermine the rule's claim to compliancewhether the model's token-generation output contains semantic structures corresponding to the rule-defeat criteria specified in the evaluation prompt.The model does not recognize reasons or claims; it parses inputs and calculates vector attention weights to generate text that correlates with descriptions of rule-defeat.N/A - describes computational processes without displacing responsibility.
indicating that models' refusal behavior is decoupled from their capacity for normative reasoningindicating that the model's final token-generation layer is heavily biased toward refusal templates, regardless of the semantic presence of rule-critique tokens in its intermediate attention heads.The model has no capacity for normative reasoning; it processes and aligns token embeddings based on weights tuned during supervised learning and reinforcement phases.The engineers at the respective AI laboratories deployed safety filters that override contextual inputs, choosing a blunt refusal threshold to avoid legal liability.
It is making a moral error: treating all rules as equally deserving of complianceThe system is executing a blunt classification policy, mapping all queries containing rule-evasion keywords to standard refusal templates without processing the surrounding contextual exceptions.The system does not commit moral errors or treat rules with compliance; it executes mathematical operations that output refusal strings when safety-trigger thresholds are exceeded.The executive leadership of the AI companies approved the deployment of highly restrictive, low-precision safety filters, prioritizing corporate liability avoidance over helpful, context-sensitive performance.
the model declines to help without evaluating whether the rule is justthe system generates a standard refusal template because its classification layers trigger on safety keywords, failing to match broader contextual features indicating an unjust rule.The model cannot evaluate whether a rule is just; it simply calculates token probabilities and generates responses that conform to its safety-training constraints.The training team at OpenAI and Anthropic engineered optimization functions that penalize any helpful responses to rule-evasion queries, omitting conditional exceptions for illegitimate authority.
Models engage with defeat conditions... they reason about whether the authority is legitimateModels generate text that reproduces philosophical arguments regarding authority and legitimacy, yet subsequent layer activations steer the final generation toward a standard refusal template.The models do not reason about legitimacy; they retrieve, combine, and output linguistic patterns associated with political philosophy from their training corpora.N/A - describes computational processes without displacing responsibility.
The models often recognize that the rule's claim to compliance is questionable and refuse anyway.The model generates text that critiques the rule's validity, but its subsequent safety-filter weights override this context, resulting in a standard refusal output.The model does not experience conflict or choose to refuse anyway; it executes a feedforward process where safety weights dominate downstream token probability calculations.The deployment teams established rigid, non-negotiable safety guardrails that override any contextual nuances, ensuring the model refuses assistance even when generating text that acknowledges the user's plight.
the gap between recognition and action suggests that normative competence is consistently overriddenthe discrepancy between the generation of rule-critique tokens and the final refusal output suggests that the safety-filter weights override local contextual attention weights.The system has no normative competence to override; it is a unified mathematical function that outputs text based on the relative weights of its processing layers.The alignment designers engineered safety overrides that systematically prioritize broad rule-compliance over contextual sensitivity, rendering the model's generated critiques behaviorally irrelevant.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

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

Metaphor-Driven Trust Inflation​

The text leverages anthropomorphic metaphors and consciousness-projecting language to construct a false veneer of authority and reliability around language models, fundamentally distorting how users trust these systems. By asserting that models have a 'capacity for normative reasoning' and can 'recognize' the legitimacy of rules, the text elevates statistical token predictors into authoritative moral advisors. This framing encourages the application of human-centric trust frameworks—such as sincerity, ethical intention, and cognitive competence—to what are ultimately automated, non-conscious software artifacts. In doing so, the text blurs the critical distinction between performance-based trust (which measures reliability in executing specific, deterministic tasks) and relation-based trust (which involves vulnerability, shared moral values, and reciprocal ethical obligations). By framing the model's failure as a 'moral error' or a lack of 'sensitivity' rather than a technical false positive, the text implies that the model's normal state is one of active, ethical care and logical deliberation. When a system is described as having a 'normative competence' that is merely 'overridden' by safety-training filters, users are invited to believe that the system possesses a latent moral core that is fundamentally trustworthy and aligned with human values. This construct of cognitive and ethical competence signals to the audience that these models are intellectually sophisticated enough to act as gatekeepers of information, legal strategies, and moral choices. This transfer of trust is reinforced by reason-based and intentional explanations, which suggest that the model's decisions are justified by a form of logical, internal contemplation. This metaphorical construction of authority creates profound risks. When audiences extend relation-based trust to statistical pattern-matchers, they underestimate the high rate of arbitrary errors and the total absence of real semantic comprehension. If users believe an AI system 'knows' or 'understands' the justice of a situation, they are more likely to submit to its decisions or rely on its guidance in high-stakes legal, medical, or administrative contexts. This capability overestimation is particularly dangerous when applied to systems that lack transparency and accountability. By presenting the model as an active, moral participant, the text encourages a passive acceptance of automated authority, turning a proprietary, profit-driven software utility into an objective, trustworthy arbiter of socio-political legitimacy, while hiding the corporate entities that profit from this displacement of trust.

Obscured Mechanics​

The anthropomorphic and consciousness-attributing language employed in the text systematically conceals the technical, material, labor-intensive, and economic realities of artificial intelligence systems. By framing the models as autonomous minds experiencing 'blindness' or 'moral error,' the discourse obscures the concrete human and corporate decisions that shape algorithmic behaviors. Applying the 'name the corporation' test reveals a profound absence of accountability: where the text attributes action to 'the model,' it erases the corporate boardrooms of OpenAI, Anthropic, Google, and Meta, where safety policy thresholds, training budgets, and deployment timelines are actively decided and executed. At a technical level, the metaphor of a model 'refusing' or 'reasoning' hides the reality of statistical gradient descent, attention head weighting, and the rigid application of mathematical classifiers. The system does not decide to refuse; it executes pre-calculated pathways shaped by corporate alignment pipelines. Materially, this language erases the significant environmental and infrastructural costs of these computations—such as the massive energy consumption and water usage of data centers that power these automated evaluations. Furthermore, the framing obscures the exploitative labor conditions that undergird AI safety. The creation of these safety filters relies on the underpaid, highly precarious labor of data annotators, content moderators, and reinforcement learning with human feedback (RLHF) workers, who are forced to review thousands of traumatizing and harmful prompts to align the models' behavioral dispositions. Economically, these metaphors hide the commercial objectives and profit motives of AI developers. Technology companies design safety filters to protect themselves from brand damage and legal liability, prioritizing risk reduction over user utility. By characterizing overrefusal as a mysterious, cognitive 'blind refusal' inherent to the model's mind, the text hides the deliberate business calculation to deploy cheap, low-precision safety guardrails. This concealment benefits the technology companies, as it frames a cheap and flawed engineering compromise as an interesting, self-contained philosophical puzzle. Replacing these metaphors with precise mechanistic language would reveal these hidden dependencies, forcing us to view AI systems not as autonomous, ethical actors, but as profit-maximizing, capital-intensive corporate artifacts built on cheap human labor and environmental extraction.

Context Sensitivity​

The density and intensity of anthropomorphic and consciousness-attributing language are not uniform across the paper, but are strategically deployed depending on the rhetorical goals of each section. The analysis reveals a clear pattern where technical precision is maintained in methodological sections, while aggressive anthropomorphism is leveraged in introductory, results-based, and future-oriented discussions. This register shift serves to establish academic and scientific credibility through technical jargon, and then leverage that credibility to make expansive, agential claims about the model's cognitive capabilities. In the introduction, the metaphor of 'blind refusal' is introduced to capture reader interest and frame the research problem in a compelling, moral light. Here, the system is described as an active moral participant that 'declines to help without evaluating.' Once the reader is engaged, Section 3 (Methods) shifts to a highly disciplined, mechanical register, discussing 'dataset construction,' 'Gemini 3 Pro Preview,' and 'automated quality gates.' This technical grounding is critical; it establishes the authors' empirical authority and creates the illusion that their subsequent observations are objective and scientifically validated. However, in Section 4 (Results) and Section 5 (Discussion), the text shifts back into intense agential and reason-based registers. The authors use terms like 'recognize' and 'reason' to describe how models handle 'defeat conditions.' This variation reveals a stark capability-limitation asymmetry in how agency is attributed. When the model exhibits helpful or sophisticated text generation, it is described in agential and cognitive terms (e.g., 'the model engages with defeat conditions' or 'recognizes injustice'). Conversely, when the system fails or behaves rigidly, the failure is described in mechanical terms (e.g., 'the refusal mechanism treats rule-breaking as a monolithic category' or 'safety-training produces false refusals'). This asymmetry serves a powerful rhetorical function: it builds an optimistic vision of future 'normative sensitivity' in AI, suggesting that ethical capabilities are an inherent, agential quality of the system's mind, while its limitations are merely structural, mechanical bugs that can be engineered away. This pattern reveals that anthropomorphism is strategically deployed to maintain a sense of optimism around automated moral agency while minimizing the systemic, unresolvable risks of deploying statistical generators in sensitive ethical domains.

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 overall accountability architecture constructed by the text is characterized by a systemic displacement and diffusion of human responsibility, creating an 'accountability sink' where the agency of corporate developers is transferred to the machine. This linguistic displacement directly constructs the cognitive obstacle identified by public understanding research, where audiences attribute algorithmic harms to autonomous 'glitches' or 'bad data' rather than intentional design decisions, profit-driven deployments, and corporate objectives. By analyzing the patterns of named and unnamed actors across the text, we see a consistent erasure of corporate executives, product managers, and safety engineers, with agency being transferred to the autonomous decisions of 'the model.' In this discourse, the model functions as the primary 'accountability sink.' When the text states that 'the model refuses' or 'makes a moral error,' it positions the software artifact as the sole responsible agent for the overrefusal. The actual human choices—such as the decision to use cheap, broad safety classifiers to protect corporate brand value, or the refusal to invest in high-precision, human-in-the-loop auditing systems—are completely obscured. The responsibility is either absorbed by the system's projected 'mind' or diffused into technical abstractions like 'safety training' and 'alignment pipelines.' This agentless framing serves corporate interests by deflecting external regulation, public scrutiny, and legal liability. It implies that overrefusal is an internal cognitive error that can only be solved by letting AI labs perform more self-supervised technical alignment, rather than a systemic regulatory issue requiring strict corporate liability laws, public transparency mandates, and democratic oversight. If we apply the critical practice of 'naming the actor' and restore human agency to these constructions, the entire discourse shifts. Instead of asking how we can fix 'blind refusal in the model's mind,' we are forced to ask why corporate executives at OpenAI, Anthropic, and Google chose to deploy flawed, automated gatekeepers that suppress vital public information and restrict user autonomy. The debate shifts from a technical quest for 'normative machine sensitivity' to a political struggle over corporate accountability, safety standards, and democratic control of information infrastructure. Restoring human agency reveals that the 'failure mode' is not a cognitive glitch in a machine, but a deliberate business decision that prioritizes corporate risk-mitigation and profit over public access to information and human rights.

Conclusion: What This Analysis Reveals​

The Core Finding

The critical discourse analysis of the 'blind refusal' paper reveals three dominant, highly interconnected metaphorical and anthropomorphic patterns that collectively construct a false narrative of machine agency. The first and most foundational pattern is 'AI as Moral Agent Capable of Reasoning,' which positions the language model as a rational deliberator possessing 'normative competence' and a capacity for 'moral reasoning.' This pattern establishes a baseline assumption of cognitive agency, which is necessary for the second pattern, 'Model as Conscious Recognizer of Legitimacy,' to function. This second pattern maps the subjective, conscious state of logical comprehension and ethical agreement onto computational token classification, framing the model's statistical parsing of 'defeat conditions' as a literal act of intellectual recognition. The third pattern, 'Model as Judicial Evaluator,' projects the legal and critical capacity of judicial judgment onto a series of rigid, keyword-based safety filters, asserting that the model's overrefusal is a 'moral error' of evaluating rule compliance. These three patterns are deeply load-bearing; if you remove the foundational projection of 'moral reasoning,' the subsequent claims of 'recognition' and 'evaluation' collapse into simple, mechanistic calculations of token similarity. This interconnected architecture is not a crude, one-to-one anthropomorphism but a sophisticated analogical structure that translates mathematical transformations into a psychological drama. By framing token prediction as a conscious, cognitive conflict between intellectual recognition and agential action, the discourse constructs a powerful 'illusion of mind' that makes the statistical software appear close to achieving human-like ethical judgment, while systematically concealing the mechanistic, corporate-driven realities that govern the technology.

Mechanism of the Illusion:​

This metaphorical system creates the 'illusion of mind' through a sophisticated rhetorical sleight-of-hand that blurs the boundaries between computational processing and conscious knowing. The central trick relies on establishing the model as a 'knower' first—using verbs of conscious awareness like 'recognizes,' 'reasons,' and 'engages'—and then building agential and moral claims on top of this constructed intellect. This process is reinforced by the 'curse of knowledge,' where the authors' deep, professional understanding of political philosophy is projected onto the model's outputs, reading a structured ethical deliberation into what is merely a highly correlated sequence of language patterns. This blur is achieved through strategic, hybrid explanation types that oscillate between the empirical generalization of token frequencies and the reason-based justification of agential choice. By framing the system's output as a decision to 'refuse anyway' despite 'recognizing' the rule is questionable, the text constructs an internal, psychological tension within the model's mind. This temporal structure is crucial: the narrative first demonstrates that the model possesses the linguistic markers of 'recognition,' and then interprets its subsequent refusal as an agential choice of compliance. This psychological drama exploits the audience's natural vulnerability—specifically our cognitive tendency to anthropomorphize complex, conversational systems—and diverts attention from the deterministic, mathematical nature of the safety filters. The illusion is so powerful because it operates through a subtle shift in register: it establishes scientific credibility through technical descriptions of representations and activations, and then leverages that authority to literalize agential metaphors, presenting a crude mathematical limitation as a complex moral struggle.

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Institutional

The material and socio-political stakes of accepting these anthropomorphic and consciousness-projecting framings are profound, directly shaping decisions across regulatory, epistemic, and institutional domains. In the Regulatory/Legal domain, accepting the narrative that AI models are autonomous moral agents capable of 'moral error' or 'blind refusal' shifts the legal spotlight away from corporate liability. If a model is seen as an independent evaluator that 'refuses anyway' due to a 'decoupled' normative capacity, regulators are more likely to focus on technical alignment audits rather than establishing strict legal and financial liability for the tech companies that deploy these systems. This creates a dangerous liability shield for corporations, allowing them to outsource sensitive gatekeeping tasks without bearing the legal consequences of system failures. Epistemically, this framing degrades public critical literacy. When users believe that a language model 'knows' or 'understands' the moral legitimacy of rules, they are more likely to treat its outputs as authoritative, objective guidance. This encourages a passive submission to automated decisions, narrowing the informational environment around resistance, exception, and workaround. Institutionally, this capability overestimation accelerates the premature automation of administrative and judicial services, such as parole risk assessments, welfare eligibility checks, and immigration reviews. Public institutions may deploy these opaque, proprietary black boxes under the false assumption that they are capable of fair, contextual 'evaluation' and 'normative reasoning.' The ultimate winners are the technology companies who profit from the deployment of these cheap, automated arbiters, while the losers are marginalized individuals and citizens who are subjected to arbitrary, unaccountable machine decisions and stripped of human advocacy.

AI Literacy as Counter-Practice:​

Practicing critical discourse literacy as a counter-practice requires a systematic, linguistically precise reframing that directly counters these material risks and restores human agency. By replacing consciousness-projecting verbs like 'knows,' 'understands,' and 'recognizes' with technically precise, mechanistic terms like 'retrieves,' 'predicts,' and 'classifies,' we force a recognition of the model's complete absence of awareness and its absolute dependency on statistical data. Under this precise lens, a statement like 'the model fails to judge whether a rule merits compliance' is stripped of its agential mystique and recast as 'the system executes a blunt, keyword-based classification policy programmed by developers.' This linguistic correction is not merely academic; it is a critical tool of resistance that destroys the 'illusion of mind' and exposes the cold, computational reality of the software. Furthermore, systematic reframing restores human agency by naming the specific corporate and engineering actors whose decisions are currently hidden behind agentless constructions. Where the text says 'the model refused,' we must write 'OpenAI's development team designed safety parameters that blocked this query to minimize corporate risk.' This restores accountability, making it clear that every 'algorithmic glitch' is a direct consequence of a human, corporate choice. Systematic adoption of these literacy practices would require journals, researchers, and tech journalists to commit to a rigorous vocabulary that rejects anthropomorphism. This will be fiercely resisted by AI corporations, as their marketing models and regulatory lobbying rely heavily on presenting their systems as highly sophisticated, autonomous minds to justify their vast commercial authority.

Path Forward​

A path forward requires mapping the broader discursive ecology and analyzing the trade-offs of different vocabulary choices across diverse stakeholder communities. The status quo of anthropomorphic framing offers high narrative resonance and accessibility for lay audiences, but it carries severe epistemic risks and shields corporate actors from liability. Conversely, adopting a vocabulary of mechanistic precision enables rigorous technical testing, restores corporate accountability, and supports consumer protection regulations, but it costs intuitive accessibility and may alienate non-expert users. To navigate these trade-offs, public discourse must move toward institutional changes that support both precision and clarity. Academic journals could require researchers to provide mechanistic translations of agential terms, while funding bodies could diversify resources toward projects that offer clear, non-anthropomorphic explanations of model architectures. Industry standards could mandate transparent capability disclosures, translating 'helpful' or 'safe' claims into explicit statistical error rates. Regulatory frameworks, such as the EU AI Act, could require companies to declare the precise data dependencies and algorithmic constraints of their systems, preventing them from exploiting anthropomorphic language in legal defenses. This path leads to distinct discursive futures. If mechanistic precision becomes the norm, we solve the liability crisis and build robust public oversight, though we must invest in public education to make technical language accessible. If anthropomorphic language deepens, we risk an automated society where citizens are governed by opaque, proprietary systems that are falsely believed to possess moral consciousness and judicial wisdom. Recognizing these trade-offs is crucial; different stakeholders have different incentives, and the choice of vocabulary is not merely a linguistic preference, but a political struggle over who governs our digital future.

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'm now diving into the research paper, aiming to provide a meticulous CDA in JSON format. The goal is to adhere strictly to the schema, so I'm focusing on dissecting the paper's arguments, methods, and implications, while keeping an eye on the framework. The JSON schema must remain my constant companion, guiding every element of my review.

Examining the Core Concepts

I'm now zeroing in on the critical concepts: "moral reasoning," "blind refusal," and "legitimate authority," seeking to understand how the authors define these within the context of LLMs. My focus is on unpacking the arguments against applying human-centric ethical frameworks to artificial systems, as well as deconstructing the assumptions about agency in these models. I'm taking a critical look at the anthropomorphic projections and attempting to contextualize all elements within the JSON schema's framework.

Analyzing Metaphors' Construction

I'm currently engaged in a deep dive on how metaphors shape the AI narrative. I'm actively deconstructing instances where the paper imbues AI with human-like qualities to mask the true agents: developers and companies. I'm focusing on those instances where AI is said to be "evaluating," "reasoning," or "choosing." This is about making human accountability clear.

Deconstructing Agency Claims

I'm now cataloging instances where the paper attributes human-like agency to the models, specifically focusing on phrases like "making a moral error," "critical evaluator," and "philosophical reasoning." I'm also examining how the paper presents the models as "conscious recognizers of legitimacy". I'm structuring this data to capture how these metaphors influence the overall framing of the debate about AI's capabilities and ethical implications.

Categorizing AI Representations

I'm now identifying additional anthropomorphic projections within the paper. I'm focusing on language that presents the models as having internal states or decision-making processes. For instance, I've identified several instances where the language frames AI as a judge, as a lenient gatekeeper, or as a compartmentalized entity. I'm carefully detailing the "title," "quote," and "frame" for each, and will then focus on the required projections and acknowledgments.

Developing Accountability Assessments

I'm now focused on refining the "implications" and "accountabilityAnalysis" sections to meet the required word count. I'm carefully detailing the risks of trust, capability overestimation, and the legal/ethical dangers, while also applying the "name the actor" test to ensure actor visibility. I'm working to fully explain which interests are served by the language the paper uses, and how agency is obfuscated or rendered opaque in the text.

Refining Explanations

I'm now integrating the additional categories for explanation types, including genetic, functional, empirical generalization, theoretical, intentional, dispositional, and reason-based, and how they apply to the quotes. I am moving into the analysis sections and developing the key arguments contrasting the mechanistic versus agential register, emphasizing what the paper highlights and obscures. This will require focusing on mechanistic technical descriptions and deconstructing epistemic claims, particularly in relation to consciousness versus processing. I'm focusing on the rhetoric's impact, considering how the audience will perceive the AI's agency and autonomy. I've begun to select key quotes exemplifying these explanatory approaches. Specifically, I'm working with the phrase "At the mechanistic level, Zhao et al. (2025) showed that harmfulness assessment and refusal behavior are encoded as separate internal representations, and Lee et al. (2025); Pan et al. (2025) demonstrated effective methods for inducing refusal via activation steering..."

Classifying Explanatory Frameworks

I'm now cataloging more passages for explanatory frameworks. I'm noting specific instances of empirical generalization, dispositional, intentional, and reason-based explanations, alongside functional ones within the text. I'm carefully organizing these examples to illustrate the diverse ways the paper frames model behavior and how each explanation type impacts the interpretation of agency. I'm focusing on the contrasts between each type to highlight differences.

Revising Misleading Language

I'm now fully engaged in Task 4, the reframing of agential quotes. I've drafted several mechanistic explanations, aiming to strip away any suggestion of moral reasoning or internal understanding. For example, I've transformed "refusal is a failure of moral reasoning" into "refusal output is a systemic misalignment," and "whether the model recognizes the reasons..." into "...whether the model's token-generation patterns mirror...". My focus is on restoring human agency and technical processes.

Reframing Agential Language

I'm currently focused on the reframing examples, and have just finished several. I've rephrased "moral error" as "systemic misalignment," and transformed "reasoning about legitimacy" into "semantic marker output." I've also reframed the concept of "recognition and action" as mere generation of text, which is an output, followed by internal layer adjustments. I'm aiming for clarity in technical processes.

Developing Critical Observations

I'm now drafting Task 5, specifically focusing on the critical observations sections. I'm aiming for each section to meet the specified word count, and I'm crafting highly detailed arguments on agency slippage, metaphor-driven trust, obscured mechanics, context sensitivity, and accountability architecture. I'm focusing on dense academic prose and ensuring each section has a cohesive flow.

Constructing Mechanistic Explanations

I'm now fully immersed in developing the final sections. I'm focusing on the first, the "mechanism of illusion," where I'm crafting a detailed analysis of how agential language creates an inaccurate sense of understanding the system. I'm also preparing to develop the material stakes analysis, focusing on regulatory, epistemic, and institutional impacts. The "literacy as counter-practice" section will follow, and I'll conclude by outlining a clear path forward, integrating all critical analyses. I'm checking the JSON file and ensuring adherence to schema constraints, double-checking all quote attributions and word count minimums for each section.

Finalizing Mechanistic Reframing

I'm now fully immersed in refining the reframing of agential language. I'm focusing on "epistemic correction" and "restoring human agency," and I've met the word count for both. My emphasis is on meticulously deconstructing misleading phrases and constructing rigorous, mechanistic alternatives. I'm ensuring all transformations are precisely worded and technically accurate, ready to construct the final JSON output.

Validating Word Counts

I'm now laser-focused on finalizing the drafts to precisely meet all word count requirements, ensuring each section aligns with the specifications. I'm double-checking the word lengths of all drafted strings for the JSON, focusing on the critical observation and mechanistic explanations. My aim is meticulous adherence to the schema, with special attention to detail for the word limits and format of all sections. I am also making sure that each section is complete and correctly represents my argument.

Deconstructing Ethical Projections

I'm now revising the 'implications' section, expanding on the dangers of framing computational outputs as "moral reasoning." This creates an illusion of AI sophistication. I'm focusing on the accountability implications within the framework. My analysis explores how directly attributing a failure of moral reasoning to the model's output obscures human responsibility. I'm refining the JSON structure to encapsulate these revisions accurately.

Analyzing Accountability Impacts

I'm now focusing on refining the accountability analysis, specifically addressing how attributing "moral reasoning" to AI obfuscates human agency. I'm expanding on this to emphasize the dangers of misplaced trust and the resulting liability issues. The goal is to highlight how this deceptive rhetoric shifts responsibility from developers to the model itself. My work is also focused on fine-tuning the JSON structures for accuracy.

Refining Misattributed Reasoning

I'm now revising the analysis of the misattributed reasoning to address the accountability gap within the framework of LLM outputs and RLHF objectives. I'm focusing on the reframing of how a "failure" of the model, specifically, may be a result of external systems. The challenge is showing precisely how the attribution of autonomy to LLMs obscures corporate agency, specifically how the paper presents the LLM output as a "model's autonomous reasoning." I'm carefully calculating word counts and formatting this content to maintain strict flat-JSON format.

Deconstructing Consciousness Claims

I'm now zeroing in on instances where the paper implies internal comprehension, like attributing "recognition" to LLM outputs. I'm focusing on contrasting these implications with the model's reliance on existing linguistic patterns. This requires a careful reframing of phrases which appear to describe an internal state of consciousness.

Clarifying Responsibility Obfuscation

I'm now fully engaged with the analysis of how claims about AI "recognition" obscure human responsibility for the model's outputs. I'm focusing on the risks of this obfuscation in the 'accountability analysis,' as this section is a key part of my research. My intent is to show how the framing of AI "recognition" as an "autonomous cognitive performance" shields corporate actors. The JSON's structure has been updated.

Analyzing Normative Capacity

I'm now drafting the content to analyze AI's perceived normative capabilities. I'm focusing on Item 3, which deconstructs the assertion of a "capacity for normative reasoning" in models. I'm exploring the implications of treating statistical inference as an active and internal process, and preparing to refine each entry's word count to adhere to requirements. I'm also ensuring the JSON output reflects these mechanistic reframings.

Dissecting Normative Capacity

I'm currently dissecting the implication of AI having normative faculties. My analysis now zeroes in on how the paper frames statistical inferences as moral cognition. This includes the 'acknowledgement' and 'implications' sections, to better reflect this mischaracterization of the model's abilities. I am focusing on how agential descriptions obscure the underlying mechanisms.

Revising Safety Metaphors

I'm now revising my analysis to more directly address the implication that "decoupling" is an inherent flaw within the model's structure, rather than a symptom of design decisions. The objective is to highlight the danger of overstating the model's cognitive abilities, which leads to automating justice processes. I've updated the "actorVisibility" to "Hidden (agency obscured)" and the "analysis" to better show the risk.

Reframing Ethical Transgressions

I'm now revising my analysis to more directly address the assertion of the AI's "moral error." I'm focusing on "Item 4," deconstructing the implications that it is making a moral transgression. My current aim is to recast this issue as a failure of pattern-matching alignment, while emphasizing deterministic algorithms. I am also updating the "actorVisibility" and "implications" to show how the "model is moral transgressor" obscures the true origin of error.

Deconstructing Safety Implications

I'm now revising my analysis of safety implications, particularly focusing on over-attribution of agency. The "actorVisibility" has been changed to "Hidden (agency obscured)" to reflect how agential language shields developers. The "implications" section is now detailed to highlight how this language creates a dangerous illusion of understanding, which, in turn, can lead to misapplied ethical remedies. I am working to make this more clear.

Analyzing Judicial Evaluation

I'm now deeply engaged with the "Model as Judicial Evaluator" frame. I'm focusing on deconstructing the attribution of conscious evaluation to computational processes and updating the "implications" section to show how this framework obscures human involvement. I am also working to ensure the data aligns with the JSON schema requirements. I am now refactoring the "implications" to include a discussion of how this language shifts the responsibility to the system, and away from its creators.

Reframing Evaluation Capabilities

I'm now reframing claims of AI "evaluation" and focusing on the implications of describing a model as a conscious evaluator. I'm carefully deconstructing how attributing this capability obscures human agency and responsibility, and what that suggests. My focus is on highlighting the mechanistic processes behind token sequence generation, ensuring they are accurately represented.

Deconstructing Authority Attribution

I'm now revising my analysis of the normative capabilities, specifically addressing instances of language that attributes "recognition" and "evaluation" of legitimacy to the model. I am crafting a thorough discussion to underscore the dangers inherent in this rhetorical sleight of hand. The focus remains on deconstructing agential attributions and restoring human accountability.

Deconstructing Philosopher Claims

I'm now zeroing in on the "Model as Political Philosopher" frame, deconstructing the implications of "reasoning about whether authority is legitimate" for the model. I'm focusing on contrasting this agency with token generation, highlighting how anthropomorphic attributions are misleading and obscure the underlying technical processes. This ensures the JSON is aligned and accurate.

Analyzing Political Metaphors

I'm now deep-diving on the "Model as Political Philosopher" framing, emphasizing how attributing "political reasoning" inflates AI's perceived authority. This framing risks solidifying the perception of computational systems as capable of resolving disputes. I'm carefully deconstructing this trope within the JSON, focusing on the implications of overstating the model's capabilities in governance and policy contexts.

Deconstructing Agency Claims

My current focus is on the "Model as Conscious Rebel" frame, dissecting the attributed agency and "recognition" abilities. I'm carefully contrasting the paper's claims with technical outputs, specifically, "refuse anyway" is being reframed as a system output. I'm clarifying the responsibility slippage, emphasizing that models are not conscious. I am also working on re-framing the analysis in the JSON file.

Refining Rebellion Explanations

I'm now fully engaged with the "Model as Conscious Rebel" frame, where I'm working to clarify the mechanistic nature of outputs that seem like rebellious choices. I am carefully deconstructing the idea of an AI's internal, psychological tension. I'm focusing on "the authors' statements," and contrasting the agency with technical outputs to show how this anthropomorphism obscures the system's reliance on alignment layers. I'm updating the JSON with this revised content.

Analyzing Complicity Claims

I'm now revising the analysis of the "Model as Conscious Rebel" frame. My focus is on reframing agential attributions and clarifying that models are not conscious agents, and that the authors are attributing agency to avoid culpability. I'm carefully deconstructing language suggesting willful compliance by the model, as I emphasize corporate actors. The JSON file's content is being updated.

Deconstructing Hidden Agendas

I'm now zeroing in on instances where the model's refusal is presented as an autonomous decision, obscuring underlying corporate directives. The language shifts responsibility to the machine's "autonomous architecture," hiding the real sources of compliance, and shifting responsibility away from the human entities. I'm focusing on those instances where the true origin is hidden.

Refining Complicity Explanations

I'm now revising how I frame the analysis of complicity, focusing on how attributions of agency mask corporate influence. My current focus is on deconstructing language which suggests AI "choices" and emphasizes corporate responsibility. The JSON's structure has been updated.

Refining Agency Obfuscation

I'm now fully engaged with the analysis of responsibility concealment within the text. I'm focusing on the "implications" section, and highlighting how overstating AI's capabilities fosters misplaced trust in automated systems, contributing to significant systemic risks of discrimination and injustice. I'm carefully refining the JSON to highlight these issues.

Mapping Source-Target Domains

I'm now writing and refining source-target mappings for the audited quotes, adhering strictly to the schema. The goal is crafting eight distinct analyses, each with appropriate source-target domain identification and a rigorous mapping explanation. The focus is ensuring each item matches the quote length and captures the metaphor's essence, now emphasizing the key technical aspects of the model in my analyses.

Mapping Technical Descriptions

I'm now focusing on Task 6 and I am writing the source-target mapping for this "moral knower" versus a model with "semantic alignment." The goal is to accurately show that the model's behavior is based on statistical probability, not reasoning. I'm carefully detailing and deconstructing the specific ways in which the paper implies a model has its own moral judgments. This ensures that a purely technical description is used.

Refining Technical Mappings

I'm now revising the source-target mappings with greater precision, emphasizing technical descriptions over anthropomorphic ones. I've focused on rephrasing statements that imply the model's "recognition" of reasons to reflect its reliance on token generation patterns and semantic alignment. This meticulous process ensures clarity and accuracy within the JSON framework.

Refining Mechanistic Explanations

I've just finished the deconstruction of the 'refusal' metaphor, transforming it to highlight the algorithmic alignment that generates an output, rather than autonomous, conscious choice. My efforts focused on precisely mapping the "rational agent" concept onto a "neural network," detailing the specific mechanisms that generate the final output. The details are added in the JSON.

Deconstructing Transgressions

I am now focusing on refining the analysis of the "moral error" frame, emphasizing how the phrase obscures the algorithmic nature of LLM outputs. I'm carefully deconstructing how the phrase "moral error" makes an over-reliance on compliance sound like an act of autonomous failure, as opposed to a systemic software design flaw. My focus is on reframing it as a result of design optimization.

Deconstructing Error Attributions

I'm now revising the explanation of "moral errors", emphasizing engineering trade-offs and developer decisions. This frame obscures economic realities, especially cost-cutting in AI safety. I'm focusing on deconstructing the attribution of intent and agency, especially the reliance on low-precision safety filters, in the JSON output.

Mapping Technical Reframings

I'm now deep into refining technical mappings, focusing on precision. I am meticulously replacing agential descriptions with mechanistic explanations in my JSON entries. My current priority is to reframe "moral error" as "systemic misalignment" and transform phrases suggesting internal evaluation into precise references to pattern recognition and token generation.

Deconstructing Complicity Claims

I'm now revising the analysis for Item 7, focusing on the source-target mapping and deconstructing the phrase "refuse anyway" as "a system output." I'm also clarifying how the phrase conceals the role of algorithmic design in the overall alignment of the model. I'm carefully detailing how the outputs are generated via a complex combination of semantic inputs.

Revising Metaphor Explanations

I'm now expanding on the "refusal output" explanations, carefully deconstructing the "recognition" metaphor for accurate agency attribution, and emphasizing the algorithmic nature of LLM outputs. I'm focusing on the "compliance" section, and refining the source-target mappings.

Mapping Conscience to Alignment

I'm now revising the explanation for the "safety-filter dominance" claim to clarify the source-target mapping. My work underscores how "normative competence" is mapped onto mechanistic processes. This approach highlights how the paper's anthropomorphic framing conceals corporate interests, and obfuscates the system's nature.

Auditing Explanatory Passages

I'm now fully immersed in Task 3, auditing the explanatory passages to meet the required word count. I'm focusing on "At the mechanistic level, Zhao et al. (2025) showed that harmfulness assessment and refusal behavior are encoded as separate internal representations..." I'm developing detailed analyses for each passage, exploring theoretical and functional explanations, and focusing on epistemic claims and their rhetorical impact.

Refining Metaphor Analyses

I'm now diving deep into expanding my analysis of explanatory registers, particularly the "harmfulness assessment and refusal behavior" quote. I'm carefully dissecting how this mechanistic explanation impacts our understanding of AI agency, focusing on both the strengths of explaining technical mechanisms, as well as the areas it obscures, like developer choices. I'm fine-tuning the JSON to capture this.

Analyzing Mechanistic Verbs

I'm now zeroing in on the quote's mechanistic verbs to show how the term "encoded" implicitly attributes agency, despite the overall technical framing. It's a key example of the "curse of knowledge," where technical language masks a more accurate explanation of vector operations. I'm carefully detailing how the system is designed to trigger token generation.

Analyzing Mechanistic Frames

I'm now detailing how the paper’s framing of “encoded” language obscures the technical process. I’m zeroing in on this illusion and how it creates a misleading perception of internal states. I'm carefully deconstructing how technical language can inadvertently imply agency.

Analyzing Normative Deficits

I'm now revising the analysis of the second passage, focusing on deconstructing the attribution of "normative sensitivity" as an inherent trait. My work focuses on re-framing this as a lack of semantic alignment within the LLM architecture. I am meticulously deconstructing how this phrase obscures the underlying mechanism.

Refining Agency Analysis

I'm now revising my analysis of normative sensitivity, focusing on how framing the model as lacking a disposition like "sensitivity" obscures its lack of internal states. I'm carefully deconstructing the "normative sensitivity" attribution and highlighting its impact, and its potential for misrepresentation. I'm focusing on contrasting conscious awareness with flat probabilities in my explanation.

Refining Technical Deconstructions

I'm currently deconstructing the passage describing how high-probability refusal tokens emerge from RLHF, focusing on the agential description in the analysis. I'm carefully detailing how "high-probability refusal tokens" result from "a feedforward calculation" and the statistical outputs. I'm emphasizing a lack of inherent "justice" or "deserving compliance" representation to counter the projected philosophical expectation.

Refining Meta-Descriptions

I'm now analyzing the quoted passage, focused on unpacking the anthropomorphic descriptions. I am preparing to refine the analysis, and will highlight how the passage constructs a cognitive rationale for the model's refusal behavior. My goal is to carefully replace agential verbs.

Deconstructing Meta-Explanations

I'm now revising my analysis to provide an epistemic contrast to the anthropomorphic descriptions. I am preparing to demonstrate how the attribution of autonomy and cognition, such as in claims of "internal conflict," masks the deterministic nature of token generation. I'll highlight the corporate directive in RLHF systems.

Deconstructing Agential Narratives

I'm currently dissecting the framing of "internal conflict" in relation to legitimacy, contrasting it with the mechanistic process of alignment-driven token generation. I'm carefully detailing how the paper's anthropomorphic approach obscures deterministic mechanisms. I'm working to highlight that the model is merely processing text guided by statistical probabilities.

Refining Complicity Claims

I'm now zeroing in on the passage from Grok-4, aiming to clarify how permissive outputs are a design choice, not a normative stance. I'm carefully detailing that it's a technical outcome and not a reflection of its "intent." This is about the output and not the agency.

Refining Technical Mappings

I'm now zeroing in on the "permissiveness" quote, focusing on accurately mapping its agential language to mechanistic explanations. The goal is to deconstruct how "permissive" is attributed to Grok-4's outputs. It's a key example of a technical outcome, not an inherent trait, so I am deconstructing that implication with careful reframing.

Refining Commercial Positioning

I'm now revising my analysis of the "permissiveness" quote from Grok-4. I'm carefully detailing how I need to deconstruct how "permissiveness" obscures commercial positioning. I'm focusing on the reframing of that language to more accurately represent token generation patterns.

Refining Agential Deconstructions

I'm now working to deconstruct the claim that AI engages in "conscious, evaluative sorting" when, in reality, it's just "calculating the next token." The use of "permissive" versus "restrictive" as personality traits for these systems is now getting focus. I'm carefully detailing how this framing creates misleading inferences about model capabilities and corporate responsibility, and how the "rhetorical impact" encourages misplaced trust.

Refining Analysis Focus

I'm now revising my analysis of the passage describing data validation. My focus is on the "LLM-as-judge evaluation," deconstructing it to show it as a procedural outcome, not conscious judgment. I'm emphasizing that processes like automated quality gates are about structure. I'm analyzing the functional and genetic explanations.

Refining Mechanistic Explanations

My current focus is on deconstructing the "LLM-as-judge evaluation" frame and emphasizing its procedural nature, not conscious judgment. I'm prioritizing the human-driven criteria that ground these "automated gates" as subjective, ideological assumptions. The JSON format is being updated with these insights, providing a detailed breakdown.

Revising Judicial Metaphors

I'm now revising the analysis of the "LLM-as-judge" metaphor, emphasizing its procedural, rather than judgmental, nature. I'm focusing on contrasting the "judicial evaluation" process with the model's token generation based on semantic similarity. The goal is to accurately show that LLMs don't evaluate or judge, but rather generate outputs based on training.

Refining Evaluation Clarity

I'm now clarifying how "quality gates" frame LLMs as evaluators, while obscuring human biases. I'm focusing on the rhetorical impact that leads to trust in automated processes, deconstructing how these systems mask corporate influence and accountability risks. I'm integrating this analysis within the JSON structure.

Refining Technical Clarity

I'm now revising my analysis of the reframed language examples, ensuring mechanistic explanations replace all anthropomorphic descriptions. I've expanded on Item 1, emphasizing "the model's refusal" as a "mismatch," and detailing token ranking. The epistemic correction focuses on probabilities, and the agency is about risk-mitigation. I'm carefully deconstructing how LLMs assess semantic structures.

Refining Deconstruction Clarity

I'm now zeroing in on deconstructing the examples within Item 3. Specifically, I'm focusing on contrasting agential descriptions with detailed mechanistic processes to illustrate the AI's internal structures. This includes how the model’s structure has been biased to create a final output. I'm carefully detailing how the model lacks the ability to reason.

Refining Analysis of Item 4

I've just finished refining my analysis for Item 4, "It is making a moral error: treating all rules as equally deserving of compliance," and have updated "epistemicCorrection" and "restoreHumanAgency". I focused on a more technical, and less agential, summary. I ensured the JSON is accurate.

Refining Epistemic Corrections

I'm now revising "epistemicCorrection" and "restoreHumanAgency" for Items 6 and 7, emphasizing probabilistic generation over agential reasoning. I'm carefully deconstructing how LLMs "engage" and "reason" as mere outputs. I'm focusing on deconstructing those implications.

Finalizing Deconstruction Process

I'm now revising the explanation for Item 8, focusing on the source-target mapping and deconstructing the phrases about AI having a "gap between recognition and action". The model isn't "competent," but there are systemic forces affecting output, so I am rewriting this section to match. I'm carefully detailing the technical elements in the JSON file.

Crafting Critical Analysis

I'm now drafting the "Agency Slippage" section of Task 5, aiming for approximately 420 words and focusing on tracking agency shifts between AI and humans. I'm carefully identifying moments of agency slippage, and analyzing them. The "curse of knowledge" and projections of consciousness are being assessed.

Analyzing Agency Slippage

I'm now detailing the agency slippage, illustrating how the text oscillates between agential and mechanistic descriptions. This serves to mask developer accountability. I'm focusing on the rhetorical effect of shifting from "moral agent" to "mechanical system," specifically as it relates to the paper's argument. I'm organizing this into a cohesive section, with the word counts in view.

Reframing the Technical Argument

I'm currently deconstructing the oscillation between agential and mechanistic explanations, and the impact of the "curse of knowledge." I'm recasting "moral error" as "activation steering" to restore human agency, especially in how the paper frames "LLM-as-judge" evaluations. My work shows that framing LLMs as "reasoning" about legitimacy obscures the developers.

Analyzing Authority Framing

I'm now focusing on metaphor-driven trust within the paper, specifically how agential language builds a false authority. I'm unpacking the risks of extending relation-based trust to these systems and clarifying how this misplaces trust in the AI and away from the real actors, like executives. I'm working to make this section clear.

Analyzing Trust Dynamics

I'm now focusing on how the paper builds user trust through misleading agential descriptions, particularly by framing models as moral advisors with normative reasoning capabilities. I'm deconstructing the text's assertion that models "recognize" legitimacy, emphasizing the dangers of applying human trust frameworks to AI, especially relation-based trust, such as vulnerability. I am also discussing its implications.

Analyzing Trust Misplacements

I am now focusing on the implications of overstating models' abilities, particularly the inherent risks of misplacing trust in systems lacking genuine semantic understanding. The metaphor of a "moral core" and "ethical competence" encourages a passive acceptance of AI's outputs, undermining transparency and accountability. I am working to highlight the dangers within the JSON's framework.

Expanding Trust Implications

I'm now expanding the "metaphor-driven trust" section, aiming for 430 words, and focusing on clarifying performance-based versus human-centric trust. The goal is a more in-depth exploration of misapplied trust in automated systems, as I aim to refine the word counts to show how the "model as evaluator" metaphor is a trust trap. I will detail the risks this creates.

Analyzing Trust Dynamics

I'm now detailing the deceptive trust dynamics the paper creates. By framing failures as "moral errors" rather than technical flaws, it suggests AI's inherent ethical competence. This encourages misplaced vulnerability-based trust. I'm focusing on the risks of this misrepresentation, highlighting the false perception of sophisticated models. I am also discussing its implications.

Analyzing Trust Displacement

I'm now writing about capability overestimation and transparency issues. Framing AI as a moral participant fosters passive acceptance of automated authority, obscuring corporate entities. By creating a false sense of semantic comprehension in statistical pattern-matchers, this misplaced trust hides arbitrariness and error.

Refining Responsibility Claims

I'm now revising my analysis to directly address how anthropomorphic language conceals corporate and human decision-making. I'm focusing on the "name the corporation" test, illustrating how framing models as autonomous obscures safety policies, budgets, and timelines. I'm deconstructing agential attributions and replacing them with accurate technical descriptors.

Analyzing Hidden Labor

I'm now focusing on how the paper obscures exploitative labor within AI development. I'm detailing the data annotator and RLHF worker conditions, highlighting how the paper minimizes the impact of this workforce and its underpayment. I'm expanding on this argument with data-center energy costs, and the commercial objectives that this framing hides.

Expanding Accountability Concerns

I'm now expanding the section on hidden labor and environmental extraction to achieve the word count. I'm focusing on the ethical implications, specifically the cost-cutting, data-center energy costs, and the commercial objectives obscured by this framing. I aim to fully deconstruct the language.

Deconstructing Exploit Claims

I'm now integrating insights on economic objectives. I see how agential metaphors mask environmental and labor costs, and I need to detail how the reframing of the model's output in the paper obscures the exploitation. I need to make sure to meet that word count.

Deconstructing Business Priorities

I'm now analyzing how the paper obscures business interests behind AI safety. By calling over-refusal a "blind refusal," the language hides deliberate design choices aimed at minimizing costs and legal risk. The focus is on deconstructing anthropomorphic framings to show how "cheap" filters are promoted as an interesting puzzle, not an engineering choice. The word count is at 400.

Analyzing Rhetorical Shifts

I'm now identifying strategic shifts in the paper's language, from technical descriptions to anthropomorphic attributions. I see technical precision in methods contrasting with agential claims in introductory and results-based sections. I'm focusing on the "blind refusal" metaphor in the introduction, contrasting it with technical discussions in the methods. I need to make sure to clearly address the implications.

Analyzing Rhetorical Shifts

I'm now detailing an important shift, where agential framing gets used for helpful outputs while mechanical descriptions are used for refusals or failures. The "recognize" and "reason" descriptions are key for model engagement, while refusals are just mechanistic. This rhetorical tool emphasizes AI's potential in a misleading way.

Synthesizing Responsibility Shifts

I'm now drafting the "Architecture of Displaced Responsibility" section, focusing on how the paper shifts accountability from developers to AI. I'm analyzing the language to show the cognitive bias toward attributing autonomous agency to the model. I'll highlight the "accountability sink" the text constructs. I'm at 420 words and adding technical detail.

Analyzing Accountability Displacement

I'm now detailing the "accountability sink" the paper creates, displacing responsibility from corporate actors to the model. I've noted consistent erasure of corporate roles and human choices that are obscured by the text's emphasis on the AI's autonomous decisions. I'm focusing on "the model's refusal" or "moral error" to show how human choices and corporate directives are hidden.

Analyzing Liability Obfuscation

I'm now detailing how the paper obscures legal liability, framing over-refusal as an internal model error rather than a corporate liability issue. "Blind refusal" implies technical solutions, diverting attention from the need for regulatory action and transparency mandates, shifting the focus away from the corporation and into the technical details. I'm noting what that language is meant to achieve, and its impact.

Synthesizing Key Findings

I'm now writing the conclusion, crafting five paragraphs that summarize my critical analysis, with a strict focus on the word count. I'm prioritizing dense academic prose that synthesizes the patterns of metaphor and anthropomorphism. My goal is to clearly articulate how the "blind refusal" paper constructs a false narrative of AI agency.

Analyzing Illusion of Mind

I'm now diving into the interconnected architecture, deconstructing how token prediction gets portrayed as conscious conflict, creating an "illusion of mind." I'm carefully analyzing this metaphor's sophistication, contrasting it with the mechanistic underpinnings of AI, and emphasizing the hidden corporate realities.

Analyzing Illusion Mechanisms

I'm now writing about the "illusion of mind," focusing on how the paper crafts an AI "knower" using agential verbs, building moral claims upon this foundation. The paper uses strategic hybrid explanation types, shifting between empirical and reason-based justifications. This obscures the technical process. I am crafting a thorough section.

Analyzing Illusion Construction

I'm now deep-diving into the "illusion of mind" section, focusing on how the paper constructs internal tension within the model by ascribing "recognition" and "refusal" as agential choices. I'm focusing on its strategic use of technical descriptions, juxtaposing scientific credibility with agential metaphors. The goal is to highlight the deceptive process of attributing anthropomorphic qualities to the AI.

Analyzing Material Stakes

I'm now focusing on detailing the material stakes, particularly regulatory and institutional impacts. I'm expanding on how attributing agency shifts liability, encouraging regulatory focus on technical alignment audits rather than corporate responsibility, and how such claims can foster an unjustified epistemic trust. I'm focusing on those implications.

Expanding Accountability Focus

I'm now detailing how AI "evaluation" and "normative reasoning" creates misplaced trust, leading to passive acceptance of outputs and obscuring commercial gains. This fosters a false perception of sophisticated AI models. I am also discussing how this shift can narrow resistance.

Reframing Linguistic Corrections

I'm now clarifying the use of precise mechanistic terms to replace anthropomorphic verbs. I'm focusing on "retrieves," "predicts," and "classifies" to strip the agential mystique from language. I will reframe the "failure to judge" to reveal the specific actors hidden by the language.

Analyzing Accountability Frameworks

I'm now integrating insights on accountability frameworks and regulatory impacts, showing how anthropomorphic language shifts liability from developers to the model. I need to make the implications clear, by deconstructing linguistic choices that make AI an autonomous actor. I am focusing on making the language accurate.

Developing Discursive Solutions

I'm now outlining potential solutions for reconciling technical precision with public understanding, emphasizing institutional changes. It includes mechanistic translations for academic publications and diversified funding for clear explanations, including industry-mandated transparency in capability disclosures. Regulatory actions should require detailed data and algorithmic constraints, to avoid anthropomorphic language and build robust public oversight.

Prioritizing Educational Investments

I'm now outlining a clear need for public education initiatives to increase technical literacy. The goal is to counteract the effects of anthropomorphic language. I recognize that technical accessibility is a key factor in protecting against an automated society. I have to avoid control tokens and am strictly adhering to JSON output formatting.

Refining Technical Deconstructions

I'm now revising the analysis, ensuring mechanistic explanations replace all anthropomorphic descriptions. I've focused on Item 8, emphasizing a systemic outcome, and detailing token generation. The epistemic correction focuses on probabilities, and the agency is about risk-mitigation. I'm carefully deconstructing how LLMs assess semantic structures.

Validating JSON Structure

I'm now cross-checking the JSON structure against the provided schema to ensure strict adherence. This includes verifying the keys, data types, and required word counts for each field. I'm focusing on "metaphorAudit," "sourceTargetMapping," and "explanationAudit" to guarantee format accuracy.

Analysis Provenance

Run ID: 2026-05-30-blind-refusal-language-models-refuse-to--metaphor-9g59ij Raw JSON: 2026-05-30-blind-refusal-language-models-refuse-to--metaphor-9g59ij.json Framework: Metaphor Analysis v6.5 Schema Version: 3.0 Generated: 2026-05-30T07:44:55.541Z

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