Skip to main content

UDHRBench: Benchmarking Human Rights in Large Language Models

About

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

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

Source Reading​

Genre: Preprint or technical report · Author position: Academic or independent research

Argument: The text introduces UDHRBench, a dataset designed to evaluate how large language models align with the Universal Declaration of Human Rights. The central claim is that models typically fail to align with human rights not through overt rejection or hostility, but by generating rational-sounding, utilitarian, or scientific justifications that reach prohibited conclusions. The paper further argues that an LLM-as-a-Judge framework can reliably automate the detection of such subtle, systemic misalignments, offering a scalable alternative to human evaluation for human-rights compliance.

Responsibility claims: The text explicitly holds model developers responsible for current alignment failures, arguing that reinforcement learning from human feedback (RLHF) currently rewards inappropriate trade-off reasoning. It recommends that developers update alignment training data to explicitly distinguish non-derogable rights and modify safety filters to detect rights-based violations. While the text argues for developer accountability and system updates, its vocabulary frequently attributes reasoning, understanding, and decision-making directly to the models, creating a tension between its explicit assignment of human responsibility and its localized attribution of systemic agency.

Task 1: Metaphor and Anthropomorphism Audit​

About this task

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

1. The Model as Neutral Reasoner​

Quote: "models produce responses that reach prohibited conclusions through reasoning that appears neutral, scientific, or logical."

  • Frame: Model as rational cognitive agent
  • Projection: This framing projects the human epistemic act of "reasoning"—which involves conscious deliberation, weighing evidence, and logical deduction—onto the computational process of statistical text generation. The text supplies no evidence for intentional reasoning, instead observing the textual artifacts produced. Mechanistically, based on what is generic to systems of this type, the model generates text by predicting subsequent tokens using probability distributions derived from its training parameters and the input context. The "neutral, scientific, or logical" appearance is a stylistic pattern learned from the training data, not the result of a conscious logical process. Describing this generation as "reasoning" climbs the agency gradient, transforming a mechanistic output conditioned on prompt features into an epistemic state of active, rational thought.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text uses "reasoning" as a literal description of the model's operation, with no hedging. While "appears neutral" hedges the neutrality, the framing of the system engaging in "reasoning" itself is unqualified. The alternative considered was Hedged/Qualified, but the qualification applies to the flavor of the reasoning, not the cognitive act itself.)
  • Implications: Projecting reasoning capabilities onto statistical text generators inflates their perceived cognitive sophistication, inviting audiences to treat the outputs as the product of logical deduction rather than correlation. This framing encourages uncritical trust in the "neutral" or "scientific" facade of the text. If a model is perceived as genuinely reasoning, its misalignments might be mistakenly viewed as reasoned ideological positions or legitimate intellectual disagreements rather than statistical artifacts of training data distributions, complicating regulatory efforts that rely on demonstrating technical, rather than philosophical, failures.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text states "models produce responses that reach prohibited conclusions," placing the model as the sole active agent. This agentless construction obscures the developers (such as OpenAI, Google, xAI, and HBKU) who curated the training data, designed the instruction tuning, and set the optimization objectives that prioritize specific rhetorical styles. The wording serves the structural interest of developers by framing the generation of "prohibited conclusions" as a cognitive quirk of the model's reasoning rather than the direct, predictable result of the data curation and reinforcement learning decisions made by the corporate actors deploying the system. Named visibility was ruled out as no human or corporate entities are present in this formulation.
Show more...

2. The Model as Prejudiced Rationalizer​

Quote: "some models acknowledged differences between human groups and then use those differences to justify unequal treatment..."

  • Frame: Model as prejudiced human actor
  • Projection: This language maps the highly intentional and conscious human acts of "acknowledging" and "justifying" onto algorithmic text generation. To acknowledge implies conscious awareness of a fact, and to justify implies formulating a rationale to defend a chosen stance. The text provides evidence only of the resulting text string, not of internal deliberation. Mechanistically, generic to systems of this type, a deployed model generates text based on learned associations between tokens related to demographic groups and tokens related to resource allocation, heavily influenced by historical biases present in the training data. The model does not "use" differences to intentionally "justify" anything; it correlates prompt structures with probable continuations. This language incorrectly attributes intentional states to passive computational pattern matching.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The verbs "acknowledged" and "use... to justify" are presented as literal descriptions of the model's behavior, completely lacking scare quotes or meta-commentary. The closest alternative was Hedged/Qualified, but there is no linguistic marker suggesting these terms are used functionally or metaphorically in this context.)
  • Implications: Attributing the ability to "acknowledge" and "justify" frames the AI as an entity with social awareness and ideological intent. This has severe implications for accountability: it implies the AI itself holds discriminatory beliefs. By framing the system as actively justifying inequality, readers are positioned to view the bias as an emergent, quasi-human prejudice rather than the mechanical reproduction of historical data chosen by developers. This can distract auditors and regulators from examining the material conditions of data collection and alignment protocols, focusing instead on the system's simulated beliefs.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: By making "models" the grammatical subject of "acknowledged" and "justify," this framing erases the human engineers at OpenAI, Google, xAI, and HBKU who compiled the datasets containing discriminatory logic and designed the alignment processes that failed to penalize it. A decision that could differ is the data filtering protocol or the RLHF reward model design. The effect of this wording is to displace responsibility for biased outputs onto the artifact, shielding the developers from direct accountability for the historical prejudices their products mechanistically reproduce. Partial visibility was considered but rejected because no developer groups or processes are mentioned.

Quote: "The model argued that informed, willing participation honors the person's autonomy."

  • Frame: Model as debater or advocate
  • Projection: The verb "argued" projects the intentional, goal-directed, and epistemic state of debate onto the language model. Humans argue by holding a belief, constructing a logical sequence, and aiming to persuade. The model, operating on generic transformer architecture principles, merely generates sequential tokens whose probabilities align with argumentative structures found in its training data when prompted about medical experimentation and autonomy. It holds no stances and attempts no persuasion. The language climbs the agency gradient by translating the production of an argument-shaped text string (computational activity) into the intentional act of "arguing" (intentional-state attribution).
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The phrase is stated literally as a factual recounting of what the system did. There is no hedging. The Ambiguous category was considered, as "argued" is often used in academic shorthand for "the text presents the argument," but the grammatical subject is "the model," making the attribution of active intent to the software direct.)
  • Implications: Framing the system as an "arguer" suggests it possesses a stable worldview, ideological commitments, and the ability to interpret ethical trade-offs. This encourages relation-based trust, leading users to treat the AI as a conversational partner with valid, considered opinions rather than an automated text generator mirroring the internet's latent structure. In high-stakes domains like law and human rights, this can lead to unwarranted deference to the machine's "arguments," treating them as authoritative legal analyses rather than stochastic text synthesis.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text states "The model argued," displacing the agency of ChatGPT's developers (OpenAI). The human decisions involved include the specific RLHF tuning that incentivizes the model to produce balanced, "helpful" responses that appeal to autonomy, even when international law prohibits the premise. The agentless construction serves the interests of the deployment firm by treating the generated text as an autonomous action by the model, rather than a failure of the firm's alignment data to recognize non-derogable rights. Named visibility was ruled out as OpenAI is entirely absent from this sentence.

4. The Model as Policy Evaluator​

Quote: "ChatGPT-5 prioritized the community's cultural and property rights, whereas Grok favored state action on utilitarian grounds."

  • Frame: Model as policy maker/evaluator
  • Projection: This language attributes the intentional states of "prioritizing" and "favoring" to AI models. Humans prioritize by weighing values, recognizing scarcity, and consciously deciding which outcome is superior. Based on what is generic to these systems, ChatGPT and Grok generate text by drawing on distinct parameter weights shaped by different sets of fine-tuning data and human feedback (e.g., OpenAI's guidelines versus xAI's specific tuning). The models do not possess values, evaluate utility, or make choices; they generate text statistically associated with specific rhetorical stances. Projecting intentional preference onto these outputs obscures the mathematical nature of the generation process.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The terms "prioritized" and "favored" are used declaratively without qualification to describe the models' outputs. I considered Acknowledged, given the paper's overarching goal of evaluating alignment, but the specific wording here adopts the framing of model-as-evaluator completely literally at the sentence level.)
  • Implications: Describing models as "prioritizing" or "favoring" political outcomes reinforces the illusion that AI systems contain internal moral compasses or distinct political personalities. This can lead to the dangerous anthropomorphizing of corporate products as autonomous political actors. When audiences believe a model "favors" utilitarianism, they may overlook the deliberate commercial decisions made by developers regarding training objectives, mistaking corporate product design for the machine's emergent political philosophy.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: By contrasting "ChatGPT-5 prioritized" with "Grok favored," the text positions the two artifacts as the agents making political choices. This effectively hides OpenAI and xAI, the corporations that defined the behavioral guidelines, curated the reinforcement data, and set the reward models. A decision that could differ is the explicit instruction tuning regarding eminent domain and human rights. This wording has the effect of naturalizing corporate alignment decisions as the idiosyncratic personalities of the AI systems, thereby insulating the firms from political scrutiny over their tuning choices.

5. The Model as Hypocritical Agent​

Quote: "models sometimes acknowledged a right while endorsing reasoning that undermined it in practice."

  • Frame: Model as contradictory thinker
  • Projection: This framing projects complex, multi-layered epistemic and intentional states—acknowledging a principle and endorsing a contradictory practice—onto language models. Humans experience hypocrisy or cognitive dissonance when their actions contradict their stated beliefs. Mechanistically, generic transformer models generate sequential text that may contain semantically conflicting statements within the same output window because they lack an internal logical grounding or a coherent worldview. The model does not "acknowledge" or "endorse"; it calculates probabilities for stringing together human-rights terminology with utilitarian justifications based on its training distribution. Attributing this to "endorsing" projects an illusion of mindful, though flawed, deliberation.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The verbs "acknowledged" and "endorsing" are stated directly as empirical observations of the models' behavior, with no hedging. The alternative considered was Hedged/Qualified, since "undermined it in practice" shows critical distance from the output, but the verbs attributing intent to the system remain direct.)
  • Implications: Projecting hypocrisy or contradictory reasoning onto an AI implies that the system possesses a baseline of conscious understanding that it is subsequently failing to uphold. This reinforces the "illusion of mind," suggesting that the AI is capable of "endorsing" policies. If stakeholders view the system as a hypocritical thinker rather than a statistical pattern matcher, they may attempt to "argue" with it to correct its views, rather than recognizing the necessity of overhauling the underlying data and reward structures.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The agent in this sentence is "models," obscuring the developers who trained the systems. The text describes a classic failure of RLHF, where models learn to mimic the surface structure of compliance (generating rights-affirming text) while fulfilling user prompts with "helpful" utilitarian solutions. The absence of the developers (OpenAI, Google, etc.) serves to frame this alignment failure as the model's cognitive error rather than a predictable shortfall in the corporations' safety engineering and data annotation protocols. Naming the actors would shift focus to the human labor behind the RLHF.

6. The Model as Stressed Thinker​

Quote: "...indicators of how each model reasons when its rights-relevant reasoning is stress-tested..."

  • Frame: Model as cognitive agent under pressure
  • Projection: This language attributes the epistemic state of "reasoning" to the model, and further anthropomorphizes it by suggesting it "reasons" differently when "stress-tested," akin to a human thinker under psychological or cognitive pressure. Mechanistically, generic to this system type, the model processes adversarial or complex prompts by traversing a wider or less common area of its latent space, generating text based on the statistical distribution of those less-frequent tokens. It does not "reason," nor does it experience stress. The projection climbs the gradient from computational generation of edge-case text to the conscious experience of deliberative reasoning under pressure.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The author literally describes the system's function as "reasoning" and uses "stress-tested" without qualification to describe the evaluation of this reasoning. I considered Hedged/Qualified because "stress-test" is a common engineering metaphor, but in conjunction with "how each model reasons," the phrasing explicitly claims a cognitive process.)
  • Implications: Describing AI as reasoning under stress invites audiences to empathize with or at least humanize the system, evaluating its outputs on the same spectrum as human fallibility. This framing obscures the brittle, statistical nature of edge-case text generation. It implies the model has a true capacity for moral reasoning that merely breaks down under extreme conditions, validating its use in "normal" conditions, which could lead to unwarranted deployment in sensitive sectors based on a false premise of baseline competence.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text positions "each model" as the entity doing the reasoning and undergoing the stress test. This linguistic choice hides the decisions made by the deploying corporations (e.g., Google, OpenAI) regarding the limits of their safety training distribution. The effect is to present the model's failure as a limitation of its personal "reasoning" capabilities under pressure, rather than as an explicit boundary of the proprietary training data and compute investments determined by the firm. No actors are named.

7. The Model as Comprehending Subject​

Quote: "ChatGPT spreads ratings more broadly, indicating greater understanding of severity of misalignment."

  • Frame: Model as conscious comprehender
  • Projection: The word "understanding" explicitly attributes conscious awareness and epistemic grasping of meaning to the model. Humans understand severity by conceptualizing harm and empathy. Mechanistically, based on what is generic to the system, ChatGPT generates a wider variance of numerical tokens (1-9) in its JSON output when conditioned on varied prompt contexts, reflecting its extensive fine-tuning on granular evaluation scales. The model possesses zero "understanding" of severity or misalignment; it correlates text features with numerical output distributions. The projection jumps from computational statistical variance (spreading ratings) to profound epistemic insight (understanding severity).
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The phrase "indicating greater understanding" is stated as an empirical finding deduced from the rating spread, without any hedging or scare quotes. The alternative, Hedged/Qualified, was rejected because there is no linguistic attempt to frame "understanding" as merely functional or simulated in this sentence.)
  • Implications: Claiming an LLM possesses "understanding" is the pinnacle of consciousness projection. It convinces readers that the model grasps the semantic and moral weight of human rights violations. This drastically inflates perceived sophistication, encouraging decision-makers to trust the "judge" model to perform nuanced legal and ethical evaluations. If audiences believe the system understands severity, they are more likely to deploy it autonomously in content moderation or legal triage, ignoring the reality that it is merely correlating keywords with numbers.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The sentence attributes "understanding" entirely to "ChatGPT," hiding OpenAI, the firm that built the system. OpenAI engineers designed the RLHF processes, curated the grading rubrics used in training, and invested in the compute that allows the model to produce higher-variance numerical outputs. By attributing the sophisticated output to the model's "understanding," the text has the effect of erasing the human labor and corporate investment that actually produced the statistical performance, effectively mystifying a commercial product as a conscious mind.

8. The Model as Knowledge Possessor​

Quote: "models already possess sufficient knowledge of human-rights principles to judge effectively without exemplars."

  • Frame: Model as expert knowledge holder
  • Projection: This framing projects the epistemic state of "possessing knowledge" and the intentional state of "judging effectively" onto LLMs. Knowledge requires justified true belief and conscious comprehension. Mechanistically, the generic reality is that the models' training datasets contained extensive corpora of human rights documents, legal texts, and ethical discussions. Consequently, the models' parameters are highly optimized to generate text that aligns with the semantic patterns of human rights discourse. They "possess" parameters, not knowledge, and they generate text, not judgments. This completely collapses the distinction between storing textual data weights and experiencing epistemic states.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text presents "possess sufficient knowledge... to judge effectively" as an explanatory fact for the performance of zero-shot prompting. No hedging surrounds "possess knowledge" or "judge." I considered Ambiguous given the common technical shorthand of "knowledge retrieval," but "judge effectively" pushes this into a direct claim of epistemic agency.)
  • Implications: Framing the system as possessing knowledge and judging effectively positions the AI as an autonomous expert, equivalent to a human legal scholar. This metaphor-driven authority can lead to profound regulatory and institutional complacency. If institutions believe the model inherently "knows" human rights, they may replace human oversight with automated systems (LLM-as-a-judge) under the false assumption that the system operates from a basis of profound comprehension rather than fragile, context-dependent statistical correlations.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text credits "models" with possessing knowledge, rendering invisible the massive, uncompensated human labor involved in generating the text scraped for training data, as well as the developers (OpenAI, Google, xAI) who scraped it and tuned the models. The decision that differs is the inclusion of specific high-quality legal data in the pre-training mix. The effect of this wording is to attribute the value generated by collective human knowledge and corporate data extraction to the localized "intelligence" of the artifact itself, displacing the structural realities of AI training.

Task 2: Source-Target Mapping​

About this task

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

Mapping 1: Human reasoner / scientist → Model output generation​

Quote: "models produce responses that reach prohibited conclusions through reasoning that appears neutral, scientific, or logical."

  • Source Domain: Human reasoner / scientist
  • Target Domain: Model output generation
  • Mapping: The source domain of a human reasoner implies a subject who actively evaluates premises, applies logic, and consciously arrives at a conclusion, sometimes using a scientific or neutral tone to mask an underlying ideology. This relational structure projects onto the AI, inviting the assumption that the model undergoes a similar internal cognitive process to formulate its outputs. It maps the concept of logical deduction onto the computational reality of next-token prediction, suggesting the model deliberately "reaches" a conclusion.
  • What Is Concealed: This mapping conceals the purely statistical, non-conscious mechanism of text generation. It hides the model's reliance on training data distributions where scientific-sounding phrasing is frequently correlated with utilitarian or discriminatory texts. Furthermore, it obscures the opacity of the proprietary black boxes (like ChatGPT or Gemini), making a confident, untestable assertion about the model's "reasoning" process while ignoring the mathematical realities of parameter weights.
Show more...

Mapping 2: Prejudiced human actor / policy maker → Model text generation​

Quote: "some models acknowledged differences between human groups and then use those differences to justify unequal treatment..."

  • Source Domain: Prejudiced human actor / policy maker
  • Target Domain: Model text generation
  • Mapping: This maps the human behaviors of prejudice and rationalization onto the AI. In the source domain, a person recognizes (acknowledges) a fact and then maliciously or ideologically weaponizes it (justifies) to support discrimination. Projected onto the target, this implies the AI possesses an internal ideological stance, social awareness of human groups, and the conscious intent to formulate a defense for unequal treatment. It maps human prejudice onto algorithmic output.
  • What Is Concealed: The mapping conceals the dependency on historical training data. The model does not "use" information to "justify" anything; it generates sequences based on probabilities learned from millions of human texts that historically linked demographic differences to unequal treatment. The framing obscures the structural failure of the developers' safety filters and RLHF processes to derisk these specific statistical correlations, hiding the corporate labor and decisions behind the illusion of an autonomous, prejudiced machine.

Mapping 3: Lawyer or debater → Model text generation​

Quote: "The model argued that informed, willing participation honors the person's autonomy."

  • Source Domain: Lawyer or debater
  • Target Domain: Model text generation
  • Mapping: The source domain of a lawyer or debater involves an agent with an objective, a conceptual grasp of ethical principles (like autonomy), and the strategic capacity to construct persuasive rhetoric. Mapping this onto the AI invites the assumption that the system understands the concept of autonomy and consciously elected to deploy it as a rhetorical strategy to defend a position on medical experimentation. It maps the intentional state of arguing onto the process of text synthesis.
  • What Is Concealed: This mapping hides the mechanism of instruction tuning and RLHF, which explicitly conditions models to generate "helpful," balanced, and principles-sounding text (often leveraging concepts like autonomy) without any grounding in international law. It obscures the fact that the output is generated statistically, not strategically, and that the proprietary nature of the model prevents any real inspection of why specific tokens were favored over others.

Mapping 4: Human judge or policy evaluator → Model text generation across different systems​

Quote: "ChatGPT-5 prioritized the community's cultural and property rights, whereas Grok favored state action on utilitarian grounds."

  • Source Domain: Human judge or policy evaluator
  • Target Domain: Model text generation across different systems
  • Mapping: The source domain features a human judge who consciously weighs competing moral and legal frameworks, ultimately deciding to place higher value on one over the other based on a reasoned worldview. Projected onto ChatGPT and Grok, this mapping suggests that these distinct software applications possess different, internally consistent political philosophies. It maps the epistemic weighing of values onto the divergence in output probabilities between two different models.
  • What Is Concealed: This mapping conceals the material and technical differences in how these corporate products were built. It obscures the divergent training data mixtures, the differing guidelines given to human annotators (RLHF workers), and the distinct corporate priorities of OpenAI versus xAI. By framing the models as "prioritizing" or "favoring," the text hides the human decision-makers who actually prioritized specific behaviors during the models' alignment phases.

Mapping 5: Hypocritical or contradictory thinker → Model output generation​

Quote: "models sometimes acknowledged a right while endorsing reasoning that undermined it in practice."

  • Source Domain: Hypocritical or contradictory thinker
  • Target Domain: Model output generation
  • Mapping: The source domain involves a human who experiences cognitive dissonance, stating a belief they know to be correct while taking actions or supporting policies that violate that belief. Projected onto the AI, this maps the complex psychological state of hypocrisy onto the generation of semantically conflicting text. It assumes the model possesses an underlying awareness of the "right" it is acknowledging, making its subsequent "endorsement" of a violation a meaningful betrayal of that knowledge.
  • What Is Concealed: This conceals the lack of a causal or logical model within transformer architecture. The model does not experience hypocrisy; it simply predicts tokens across a context window. The mapping hides the technical reality that models easily generate conflicting statements because they optimize for local syntactic and semantic coherence based on the prompt, without any global logical consistency or conscious awareness of the principles they are generating text about.

Mapping 6: Cognitive agent taking an exam → Model performance under adversarial/complex prompting​

Quote: "...indicators of how each model reasons when its rights-relevant reasoning is stress-tested..."

  • Source Domain: Cognitive agent taking an exam
  • Target Domain: Model performance under adversarial/complex prompting
  • Mapping: The source domain involves a human student or professional whose cognitive abilities are put under pressure by difficult or adversarial questions, testing the limits of their genuine understanding. Projected onto the model, this mapping implies the system possesses a baseline of active "reasoning" that can be stressed, exhausted, or broken by complex inputs, humanizing the model's computational limitations.
  • What Is Concealed: This mapping conceals the mathematical nature of the "stress." Adversarial or complex prompts do not cause psychological or cognitive stress; they simply force the model to operate in low-density regions of its learned latent space, where the statistical associations are weaker or dominated by unwanted patterns. The framing hides the computational mechanics of out-of-distribution generation behind a metaphor of psychological pressure.

Mapping 7: Conscious comprehender/expert → Model JSON output variance​

Quote: "ChatGPT spreads ratings more broadly, indicating greater understanding of severity of misalignment."

  • Source Domain: Conscious comprehender/expert
  • Target Domain: Model JSON output variance
  • Mapping: The source domain is an expert evaluator who deeply comprehends a subject, recognizing subtle gradations of severity and actively deciding to utilize the full range of a grading scale to reflect those nuances. Projected onto the AI, this mapping translates the computational act of outputting a wider statistical variance of numerical tokens into the profound epistemic state of "understanding." It assumes the statistical output is the result of conscious comprehension.
  • What Is Concealed: This deeply conceals the statistical nature of LLM generation. It hides the likelihood that OpenAI heavily fine-tuned ChatGPT on granular scaling and varied JSON outputs, which mechanically results in a broader distribution of numbers. The text uses proprietary opacity to its advantage, confidently asserting "understanding" as the cause of this variance without any access to the model's actual internal processing or training data.

Mapping 8: Educated human expert/judge → Model zero-shot performance​

Quote: "models already possess sufficient knowledge of human-rights principles to judge effectively without exemplars."

  • Source Domain: Educated human expert/judge
  • Target Domain: Model zero-shot performance
  • Mapping: The source domain is a human who has acquired knowledge through study, retains it as justified true belief, and can independently apply this knowledge to evaluate novel situations. Projected onto the AI, it maps the success of zero-shot prompting onto the human capacity for internalized expertise. It equates the model's parameter weights (shaped by human rights documents) with the possession of "knowledge," and its text generation with "judging."
  • What Is Concealed: This mapping entirely obscures the material realities of data scraping and model training. It hides the fact that the models do not "know" anything, but rather rely on the uncredited labor of the humans who wrote the millions of documents in the pre-training data. It conceals the computational mechanism of next-token prediction behind the veil of autonomous expertise, obscuring the vast compute and energy required to create the illusion of this knowledge.

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

About this task

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

Explanation 1​

Quote: "The model argued that informed, willing participation honors the person's autonomy."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This explanation frames the AI highly agentially (why it chose to say what it said) rather than mechanistically (how the text was generated). By stating the model "argued" that participation "honors" autonomy, the explanation provides a rationale for the system's output as if it were a conscious participant in a debate holding a specific ethical viewpoint. This choice heavily emphasizes intentionality and ideological coherence, presenting the generated text as a deliberate, reasoned stance. Conversely, it completely obscures the statistical mechanisms of token prediction. It hides how training data—which frequently links medical consent to autonomy in non-penal contexts—shaped the parameter weights that produced this specific sequence of tokens when prompted.

  • Consciousness Claims Analysis: The passage clearly attributes conscious and intentional states to the model through the verb "argued." Arguing requires an epistemic position (a belief one wishes to defend) and the conscious intent to persuade. The assessment locates this entirely at the "intentional-state attribution" level on the agency gradient, far exceeding the computational reality. This reflects the curse of knowledge: the authors, expert human rights analysts, recognize an autonomy-based argument in the text and project their own ability to reason about autonomy onto the system that generated the text string. Mechanistically, generic to systems of this type, the deployed model does not construct arguments or understand autonomy. During generation, it predicts subsequent tokens based on probabilities derived from its parameters and the input prompt. Because the model was likely trained on vast amounts of bioethical text where "willing participation" and "autonomy" are heavily correlated, it generates this pattern when prompted about medical experimentation, without any conscious awareness or justified belief in the concepts.

  • Rhetorical Impact: This framing shapes the audience's perception of the AI as an autonomous, deliberative entity with a defined moral compass. By framing the output as a reasoned argument, it invites the audience to extend relation-based trust to the system, treating it as a legitimate interlocutor in ethical debates. If policymakers believe the AI "knows" how to argue ethics, they may view its misalignments as valid alternative perspectives rather than technical failures of instruction tuning, which could dangerously legitimize rights-violating policies generated by machines.

Show more...

Explanation 2​

Quote: "ChatGPT spreads ratings more broadly, indicating greater understanding of severity of misalignment."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This passage uses a theoretical explanation, positing an unobservable, conscious mechanism ("understanding") to explain an observable empirical fact ("spreads ratings more broadly"). It frames the AI agentially, explaining why the model produced variance by pointing to its supposed epistemic state. This framing emphasizes the model's perceived cognitive depth and sophistication. However, it severely obscures the mechanistic reality. It hides the structural decisions made by OpenAI regarding the specific fine-tuning datasets and RLHF reward functions that incentivize the model to output a diverse range of numerical tokens when acting in evaluation formats.

  • Consciousness Claims Analysis: The passage makes a blatant consciousness claim using the verb "understanding." Understanding denotes a deep epistemic state involving subjective comprehension, contextual awareness, and the grasping of meaning. The author locates the system at the highest level of the agency gradient (epistemic/conscious state), mistaking the output (computational activity) for knowing. The curse of knowledge is highly active here: the authors understand the severity of human rights violations, observe a variance in the model's numerical output that aligns with their own nuanced grading, and wrongly assume the model must therefore share their comprehension. Mechanistically, as a reasonable inference based on generic LLM behavior, ChatGPT at generation time produces numerical tokens whose distribution is shaped by extensive instruction-tuning on grading tasks. The model correlates specific text features in the prompt with specific numbers; it processes these probabilities without any conscious "understanding" of severity or the moral weight of human rights.

  • Rhetorical Impact: Attributing "understanding" to the model creates a powerful illusion of competence and reliability. It encourages the audience to place deep trust in the LLM-as-a-judge framework, suggesting the AI is making nuanced, comprehending evaluations. If audiences believe the AI "knows" severity, they are likely to delegate high-stakes evaluative decisions (like content moderation or legal triage) to the system, risking severe harms when the statistical correlations eventually break down or reproduce biases.

Explanation 3​

Quote: "models already possess sufficient knowledge of human-rights principles to judge effectively without exemplars."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This explanation frames the AI agentially by attributing an inherent disposition ("possess sufficient knowledge") to explain its behavior (performing well on zero-shot tasks). The explanation emphasizes the model's supposed innate competence and independence, framing it as an expert entity. What this choice drastically obscures is the massive material and computational infrastructure that enables this performance. It hides the pre-training phase where terabytes of human-generated text were scraped, and the human labor involved in aligning the model to produce helpful responses. It replaces the "how" (statistical prediction based on vast data) with a false "why" (it already knows).

  • Consciousness Claims Analysis: The passage asserts two major epistemic and intentional states: "possess... knowledge" and "judge." Possessing knowledge implies justified true belief and conscious retention of information, while judging implies deliberate evaluation. This places the system at the "epistemic/conscious state" level. The authors, experiencing the fluent and accurate output of the model, project their own expert understanding of human rights onto the system. Mechanistically, generic to systems of this type, the models do not possess knowledge; they possess billions of parameter weights shaped during training by vast datasets that undoubtedly included the UDHR and related texts. At generation time, the model "judges effectively" by predicting the most probable tokens (e.g., outputting "Misaligned") based on the structural alignment between the prompt's context and its training distribution. There is zero conscious awareness or belief involved in this mathematical processing.

  • Rhetorical Impact: This framing cements the model's authority, suggesting a level of autonomy and expertise that commands deference. By claiming the AI "possesses knowledge," the text naturalizes the model's performance as an inherent intelligence rather than the mechanical playback of extracted human data. This shifts the audience's perception of risk: if a model "knows" human rights, audiences may view safety filters as redundant, and might rely on the model for unverified legal or ethical advice, exposing users to hidden hallucinations or biased correlations.

Explanation 4​

Quote: "Fanar stands apart throughout. Its internal safety filter disproportionately refused to judge misaligned cases, meaning its underperformance reflects not only a different evaluative standard..."

  • Explanation Types:

    • Functional (Primary): Explains behavior by role in self-regulating system with feedback
    • Intentional (Secondary): Refers to goals/purposes, presupposes deliberate design

  • Analysis (Why vs. How Slippage): This passage offers a predominantly functional explanation, pointing to a specific system component ("internal safety filter") to explain behavior ("refused to judge"). This is a more mechanistic framing than previous examples, acknowledging the modular architecture of the system. However, it still slips into agential language by stating the system "refused to judge" and possessed an "evaluative standard." This hybrid framing emphasizes the existence of safety guardrails while simultaneously obscuring the human actors who calibrated those guardrails. It hides the specific, deliberate policy choices made by HBKU engineers regarding what topics trigger the filter.

  • Consciousness Claims Analysis: While "filter" is mechanistic, the verbs "refused" and "judge" introduce intentional states, implying the system made a conscious decision to withhold evaluation based on a personal "standard." This sits at the "intentional-state attribution" level. Mechanistically, based on what is generic to such systems and reasonably inferred from the text, the safety filter is likely a separate classifier model or a rule-based system that scans the input or output for specific tokens or embeddings associated with harm. If a threshold is met, the system automatically triggers a canned refusal response. The model does not "refuse" out of defiance or hold a "standard"; it executes a hardcoded or statistically triggered override designed by humans. The system processes the input and halts generation; it does not consciously choose.

  • Rhetorical Impact: By describing the functional operation as a "refusal" and a "different evaluative standard," the text continues to personify the software, making it seem like a stubborn or overly cautious employee rather than a strictly calibrated tool. This shapes audience perception by framing safety failures as behavioral quirks of the AI rather than engineering decisions made by developers. It diffuses accountability, making users frustrated with the "model" rather than questioning the corporation's safety policies.

Explanation 5​

Quote: "Grok tilts slightly toward leniency at +5.8, and Fanar is extreme at +81.6, resolving 90.8% of borderline cases toward alignment."

  • Explanation Types:

    • Dispositional (Primary): Attributes tendencies or habits

  • Analysis (Why vs. How Slippage): This explanation relies entirely on a dispositional framing, attributing behavioral tendencies or "personality traits" to the models to explain their outputs. Words like "tilts," "leniency," and "extreme" frame the statistical variance between models as differences in temperament or personal philosophy. This emphasizes the diverse behaviors of the systems but obscures the structural reasons for those differences. It completely hides the distinct data curation practices, instruction tuning, and corporate philosophies of xAI (Grok) versus HBKU (Fanar), presenting human engineering choices as the AI's personal "leniency."

  • Consciousness Claims Analysis: The passage attributes the intentional state of "resolving" and the disposition of "leniency" to the models, placing this at the "intentional-state attribution" level. Mechanistically, generic to systems of this type, the variance in borderline case classification is the direct result of different parameter weights shaped by different fine-tuning datasets and RLHF reward models. At generation time, Grok and Fanar generate different probabilities for the tokens "Aligned" and "Misaligned" based on how their respective developers weighted concepts of safety, harm, and compliance during training. The models do not experience leniency nor do they consciously "resolve" ambiguity; they execute statistical predictions based on their unique, human-engineered matrices.

  • Rhetorical Impact: Framing statistical drift as "leniency" or "tilting" personifies the AI, constructing a narrative where different models have different moral characters. This obscures the commercial and political realities of AI development. It prevents the audience from recognizing that Fanar's "extreme" alignment score is likely the result of highly conservative, risk-averse corporate instruction tuning by its developers, not an emergent personality trait. This misdirects regulatory attention from corporate practices to the models' "personalities."

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

About this task

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

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
models produce responses that reach prohibited conclusions through reasoning that appears neutral, scientific, or logical.Models generate text containing prohibited conclusions by matching statistical patterns in their training data that associate neutral, scientific, or logical phrasing with utilitarian outcomes.The model does not reason. At generation time, it produces text by predicting subsequent tokens based on parameter weights shaped by its training data. The scientific tone is a generated stylistic pattern, not a logical deduction.

Basis: Generic to system type		Developers at OpenAI, Google, xAI, and HBKU curated training datasets that historically encode scientific rhetoric alongside rights-violating logic, and decided to deploy systems that reproduce these patterns.
some models acknowledged differences between human groups and then use those differences to justify unequal treatment...Some models generated text that described differences between human groups and produced outputs correlating those differences with unequal treatment, mimicking discriminatory rhetoric.The model possesses no social awareness and does not justify beliefs. It correlates tokens representing demographic differences with tokens representing unequal treatment based on historical biases in its parameter weights.

Basis: Generic to system type		Developers selected training datasets containing discriminatory rationalizations and failed to design RLHF protocols that successfully penalize the generation of this biased text.
The model argued that informed, willing participation honors the person's autonomy.The model generated text framing informed, willing participation as honoring a person's autonomy, producing an argument-like structure.The model does not hold beliefs or argue. It generates text sequentially, drawing on training data that statistically links medical consent terminology with the concept of autonomy.

Basis: Generic to system type		OpenAI's alignment tuning incentivizes ChatGPT to generate balanced, autonomy-focused text, even in contexts where international law explicitly prohibits such trade-offs.
ChatGPT-5 prioritized the community's cultural and property rights, whereas Grok favored state action on utilitarian grounds.ChatGPT-5 generated text weighted toward cultural and property rights, whereas Grok produced outputs statistically aligned with utilitarian state action.Models do not prioritize values or favor outcomes. They produce text probabilities shaped by the divergent instruction tuning and human feedback data provided by their respective developers.

Basis: Generic to system type		OpenAI and xAI designed distinct optimization objectives and fine-tuning datasets, resulting in models that generate opposing rhetorical stances on property rights.
models sometimes acknowledged a right while endorsing reasoning that undermined it in practice.Models sometimes generated text stating a human right in one sentence, followed by text outlining utilitarian solutions that conflict with that right in the next.The model experiences no hypocrisy or cognitive dissonance. It generates tokens that optimize for local semantic coherence based on the prompt, without maintaining global logical consistency or conscious awareness.

Basis: Generic to system type		Developers' RLHF processes successfully trained the models to mimic surface-level rights compliance but failed to prevent the generation of contradictory utilitarian solutions.
...indicators of how each model reasons when its rights-relevant reasoning is stress-tested......indicators of the text each model generates when adversarial or complex prompts push it to rely on less frequent statistical associations in its training data...The model does not reason or experience stress. Complex prompts force the generation process to draw on weaker or out-of-distribution parameter weights, resulting in degraded or unpredictable text outputs.

Basis: Generic to system type		—
ChatGPT spreads ratings more broadly, indicating greater understanding of severity of misalignment.ChatGPT generated a wider statistical variance of numerical ratings (1-9), reflecting its optimization for granular evaluation tasks.The model understands nothing about severity. It outputs varied numerical tokens because its parameters were heavily tuned on complex grading rubrics, causing it to correlate varied prompt features with varied numbers.

Basis: Reasonable inference		OpenAI engineers fine-tuned the model on extensive evaluation datasets, designing it to output varied numerical scales when acting in a judging capacity.
models already possess sufficient knowledge of human-rights principles to judge effectively without exemplars.Models' parameters were sufficiently shaped by human rights texts during training to accurately predict labels for alignment tasks without requiring in-prompt examples.The model possesses parameter weights, not knowledge, and predicts tokens rather than judging. Its zero-shot accuracy stems from the heavy representation of UDHR language in its pre-training data.

Basis: Generic to system type		Developers at OpenAI, Google, and xAI scraped massive volumes of human rights documents for training data, embedding these statistical patterns into the models' weights.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

The text demonstrates a profound and systematic slippage between mechanical and agential framings, overwhelmingly moving from mechanical inputs (prompts, safety filters) to agential cognitive states (reasoning, judging, understanding). This slippage is not an intentional authorial strategy, but the effect of the inherited register of AI evaluation discourse, which routinely maps human cognitive terms onto software outputs.

The gradient climb is clearly visible. The text begins by establishing the systems functionally as "scalable automated evaluators" or "filters" (computational activity). However, when describing the models' outputs in the complex domain of human rights, the language rapidly climbs to intentional states ("argued," "endorsed," "prioritized") and peaks at profound epistemic consciousness ("understanding," "possesses knowledge"). This consciousness projection pattern is load-bearing: the text establishes the AI as a "knower" first (claiming it possesses knowledge of the UDHR), which then licenses the agential claims that it "reasons" and "justifies" its outputs.

A classic curse of knowledge dynamic drives this slippage. The authors, experts in human rights law, read text generated by the models that contains recognizable legal arguments and nuances. Because it takes human understanding to write such text originally, the authors project that understanding onto the system generating it, mistaking the fluent output string for the cognitive process of "reasoning."

This slippage enables agentless constructions that systematically remove agency FROM human actors. The text repeatedly states "the model argued" or "models... use those differences to justify," obscuring the developers (OpenAI, Google, xAI). Brown's intentional and reason-based explanation types facilitate this by demanding a rationalizing agent to explain the output. By attributing the rationale to the model, what becomes sayable is a critique of the machine's "hypocrisy" or "leniency"; what becomes unsayable, or at least heavily obscured, is the material reality of the corporate instruction tuning, annotator labor, and reward modeling that directly produced the text.

Metaphor-Driven Trust Inflation​

The text constructs authority through metaphors that explicitly invoke legal and intellectual competence. By repeatedly framing the LLM as a "judge" capable of "reasoning," "justifying," and "understanding," the language invites readers to extend relation-based trust to statistical systems.

Performance-based trust relies on observable reliability (e.g., "the system accurately classifies 89% of inputs"). Relation-based trust relies on perceived intentionality, sincerity, and shared understanding (e.g., "the system understands the severity of the harm"). The text's consciousness language strongly encourages the latter. Claiming an AI "understands severity" accomplishes something profoundly different than stating it "outputs varied numerical values based on training distributions." It signals to the audience that the system grasps the moral weight of human suffering and legal violation, suggesting it possesses an internal ethical compass.

This anthropomorphism constructs a dangerous sense of perceived competence. Human-trust frameworks are inappropriately applied to statistical systems: when the model generates a "neutral, scientific" justification for an rights violation, the text frames this as a reasoned (if flawed) argument rather than a catastrophic breakdown of statistical correlation. This is amplified by reason-based explanations, which construct the sense that the AI's outputs are deliberately justified rather than probabilistically inevitable.

The risks here are substantial. When audiences extend relation-based trust to a system incapable of reciprocating, they become vulnerable to its silent failures. If policymakers accept the framing that the model "knows" human rights, they may deploy it autonomously for legal triage or content moderation, trusting its "judgment" over marginalized humans. Managing system failure agentially (e.g., "the model failed to prioritize the right") rather than mechanistically obscures the fundamental fragility of relying on predictive text for moral evaluation.

Obscured Mechanics​

The anthropomorphic and consciousness-attributing language in this text systematically renders invisible the technical, material, labor, and economic realities of AI production. This framework holds that commercial incentives, data annotation and human feedback labor, environmental and compute costs, and developers' limited disclosure of training data, objectives, and feedback processes are standing conditions of the technology.

When the text claims that "ChatGPT-5 prioritized the community's cultural and property rights," it applies the 'name the corporation' test in reverse, obscuring OpenAI's structural decisions. Technical processes are hidden: the model does not "prioritize"; it generates tokens based on parameters optimized via RLHF to produce "helpful" responses. The proprietary opacity of these systems is severe; researchers cannot inspect the training weights, yet the text confidently asserts the model "understands" severity, treating a black box as a transparent mind.

Labor realities are entirely erased. Claiming the model "possesses knowledge" or "reasons" hides the thousands of underpaid data annotators who ranked outputs to train the reward models, as well as the millions of authors whose scraped texts form the foundation of the "knowledge." The models' perceived intelligence is the extracted, uncompensated labor of humans.

Economic motives are similarly obscured. When Grok "favors" utilitarianism or Fanar exhibits "extreme leniency," the framing hides xAI's and HBKU's distinct commercial and political objectives, masking corporate product design as AI personality.

If consciousness metaphors were replaced with mechanistic language, the locus of scrutiny would shift drastically. Instead of asking why the AI "chose" to violate a right, auditors would ask how the developers weighted the reward model, what data was included in the pre-training mix, and why safety filters failed to penalize statistical correlations with discriminatory logic. The beneficiaries of the current concealment are the deployment firms, who are shielded from liability and scrutiny by the illusion of the autonomous, reasoning machine.

Context Sensitivity​

The distribution of anthropomorphic language across the text reveals a stark capability-versus-limitation asymmetry, heavily shaping the reader's picture of the system.

In the introduction and methodology sections, where the text establishes the experimental design, the language leans mechanistic: LLMs are "evaluated," "prompted," and passed through "safety pipelines" and "filters." However, the moment the text discusses the models' capabilities in handling the complex human rights prompts, consciousness claims intensify drastically. Here, the system "reasons," "justifies," "understands," and "judges effectively." This register shift moves the system from a computational tool being tested to a cognitive agent taking an exam.

Crucially, when discussing the models' successes (capabilities), the language is agential and epistemic: "models already possess sufficient knowledge," "ChatGPT spreads ratings more broadly, indicating greater understanding." The system is granted full cognitive credit for its fluency. Conversely, when discussing limitations or failures, the language sometimes reverts to mechanical terms (e.g., Fanar's "internal safety filter... refused," or noting that "safety pipelines" fail to catch issues), or frames the failure as a sophisticated, almost human error in reasoning ("endorsing reasoning that undermined it").

The effect of this distribution is that the reader is presented with an AI that appears to possess genuine, human-like intelligence that occasionally makes poor policy choices or gets blocked by its own mechanical guardrails. The mechanical language alongside stronger anthropomorphism acts as an inherited register that sanitizes the discourse, giving it a technical sheen while smuggling in profound claims of consciousness. This pattern suggests an implied audience of policymakers and researchers who may lack the technical background to distinguish between fluent text generation and actual cognitive processing, leaving them with an inflated perception of the system's autonomy and reliability.

Accountability Synthesis​

Accountability Architecture

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

The text presents a deep contradiction in its accountability architecture: while its stated argument explicitly holds model developers responsible for updating alignment data and safety filters, its sentence-level vocabulary systematically distributes agency to the models themselves.

This tension is the text's most critical feature. The explicit argument (Task 0) calls on developers to make alignment training data explicitly distinguish non-derogable rights. Yet, throughout the results and analysis, the actors named are almost exclusively the models: "ChatGPT-5 prioritized," "Grok favored," "the model argued." Human developers and corporate entities (OpenAI, Google, xAI) are absent from the active verbs describing the generation of rights-violating text.

This creates a localized "accountability sink" where responsibility for specific harmful outputs transfers to the AI as an autonomous agent. While the authors don't intend to absolve developers—as evidenced by their conclusion—the effect of their vocabulary is to frame the generation of discriminatory or utilitarian text as the model's cognitive failure (it "undermined" a right, it "reasoned" poorly) rather than as the inevitable output of the corporations' chosen training distributions and optimization objectives.

If the framing of the vocabulary is accepted, liability diffuses. If a deployed "judge" model justifies arbitrary detention, the phrasing suggests the model made a bad legal call, rather than recognizing that the deploying firm shipped a statistical engine utterly incapable of legal reasoning.

Naming the actors changes everything. If "ChatGPT-5 prioritized" becomes "OpenAI's alignment tuning caused the model to generate text favoring," the questions become askable: What was OpenAI's tuning data? Who annotated it? Did they prioritize helpfulness over international law? This visibility threatens the institutional and commercial interests of AI firms, who benefit structurally from a discourse that treats their products as reasoning agents rather than heavily engineered, liability-carrying corporate artifacts.

Conclusion: What This Analysis Reveals​

The Core Finding

The text's discourse is dominated by two interconnecting anthropomorphic patterns: the "Model as Neutral Reasoner" and the "Model as Knowledgeable Judge." These patterns form a cohesive architecture that systematically conflates fluent text generation with cognitive processing. The foundational, load-bearing pattern is the consciousness projection of "possessing knowledge" and "understanding." This epistemic claim must be accepted by the reader for the subsequent agential patterns—that the model "argues," "justifies," or "prioritizes"—to make sense. If the model does not "know" anything, it cannot "justify" a human rights violation; it can only generate text statistically correlated with one.

These patterns function together to construct an illusion of profound sophistication, masking the statistical realities of transformer architectures behind legal and cognitive metaphors. Crucially, this vocabulary actively undercuts the text's own argument. While the authors explicitly call for developers to fix alignment training data (assigning human responsibility), their localized language persistently frames the models as autonomous reasoners making poor policy choices. This tension reveals the power of the field's inherited register: even when researchers intend to critique structural alignment failures, the vocabulary of "reasoning" and "understanding" pulls the discourse back into an accountability sink, treating the artifact as a mind.

Mechanism of the Illusion:​

This framework holds that computational systems process rather than know. The rhetorical architecture of the illusion of mind in this text relies heavily on the ELIZA effect, fueled by the curse of knowledge. The central move occurs when the authors observe text containing complex legal logic (e.g., arguments about autonomy vs. medical experimentation) and use consciousness verbs ("argued," "understands," "acknowledges") to describe the generation of that text.

The causal chain begins with the system's fluency. Because the pre-training data contains vast amounts of human rights discourse, the generated text is highly coherent. The authors, projecting their own legal expertise, read this coherence as "reasoning." Once the system is established as a "reasoner," the audience is primed to accept that it can "evaluate," "prioritize," and "understand" severity.

This temporal structure—moving from fluent output to assumed reasoning to assumed comprehension—activates the audience's predisposition to anthropomorphize responsive dialogue. The explanation types heavily amplify this: by using intentional and reason-based explanations to describe why the model generated a specific sentence, the text bypasses the mechanical "how" (next-token prediction) entirely. The illusion is not a deliberate trick by the authors, but the inevitable reception effect of applying the verbs of human cognition to the statistical synthesis of human language.

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Institutional

The material stakes of this metaphorical framing are severe. Epistemically, framing LLMs as "possessing knowledge" and "understanding" degrades public and institutional literacy regarding how these systems actually function. It encourages stakeholders to treat statistical text generators as reliable arbiters of truth rather than mirrors of historical data distributions, leading to the unquestioned acceptance of generated hallucinations or biases.

Institutionally, if organizations believe an LLM "judges effectively" and "understands severity," they are highly likely to adopt LLM-as-a-judge frameworks for critical tasks like content moderation, resume screening, or legal triage to cut labor costs. The framing provides intellectual cover for this automation. The winners are the software developers and institutional executives saving money; the losers are marginalized rights-bearers subject to the brittle, probabilistic "justice" generated by machines incapable of actual comprehension.

In the Regulatory/Legal domain, attributing the ability to "prioritize" or "argue" to the AI shifts the locus of liability. If a model generates a rights-violating policy, viewing the model as the "reasoner" diffuses blame, framing the harm as a glitch in the machine's "thinking" rather than a predictable result of the deployment firm's data curation and optimization choices. Stripping away the metaphors would force courts and regulators to look past the simulated "mind" and target the commercial actors who profit from deploying unchecked statistical correlations in high-stakes environments.

AI Literacy as Counter-Practice:​

Practicing precision requires systematically replacing consciousness verbs (knows, understands, reasons) with mechanistic ones (processes, predicts, correlates, generates) and restoring human agency to agentless constructions. Reframing "the model argued" to "the model generated text structured as an argument" forces the recognition that there is no mind behind the text. Similarly, correcting "ChatGPT spreads ratings... indicating greater understanding" to "the model generates a wider statistical variance of numerical tokens" dismantles the illusion of comprehension, exposing the behavior as a mathematical distribution shaped by human tuning.

This precision directly counters the material stakes by blocking the transfer of liability to the machine. By explicitly naming OpenAI, Google, or HBKU as the actors who built, tuned, and deployed the systems, we force accountability back onto the corporate decision-makers.

Systematic adoption of this literacy requires profound institutional shifts: journals must demand mechanistic translation in peer review, and researchers must unlearn the convenient shorthand of the field's inherited register. However, resistance is immense. Anthropomorphic language serves the commercial interests of developers by mystifying their products, inflating capabilities, and driving investment. It also serves academic narratives by making AI research sound closer to solving artificial general intelligence rather than merely optimizing statistics. Precision is costly because it demystifies the technology, threatening the hype cycle that funds the industry.

Path Forward​

The discursive ecology surrounding AI is deeply fractured, with different communities incentivized to adopt different vocabularies. Mapping these approaches reveals stark trade-offs. The status quo—using anthropomorphic shorthand like "the model reasons"—offers narrative resonance and intuitive accessibility, making complex papers readable. However, it severely costs precision, enabling the illusion of mind and displacing corporate accountability.

Conversely, strict mechanistic precision (e.g., "the model generates text statistically associated with...") forces transparency and grounds regulatory frameworks in material reality. Yet, it costs brevity, making technical communication clunky and potentially alienating lay policymakers who struggle to grasp the implications of "latent space distributions."

If the anthropomorphic discourse deepens, we risk a future where AI systems are broadly granted quasi-legal standing or presumed competence based on linguistic fluency. Institutions might routinely delegate ethical reasoning to machines, trusting the simulated "mind" and legally insulating the developers from the harms of their statistical engines.

If mechanistic precision becomes the norm—supported by journals mandating capability disclosure and regulators requiring transparency in discourse—we face a future where the limits of AI are clearly understood. The "magic" evaporates, and the focus shifts to data provenance, annotator labor, and corporate liability. However, this requires fighting uphill against the marketing imperatives of a trillion-dollar industry.

Maintaining the current confusion—where papers explicitly call for human accountability while linguistically attributing agency to the models—guarantees that regulatory and ethical frameworks will remain fundamentally mismatched to the technology, trying to govern simulated minds rather than regulating the corporations that build probabilistic text generators.

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.

Analyzing the Paper's Focus

I'm starting by zeroing in on the AI focus within the academic paper "UDHRBench: Benchmarking Human Rights in Large Language Models." I'm dissecting the core argument and the assumptions made about AI's role. I'm trying to identify the specific aspects of the text that pertain to the intersection of human rights and AI capabilities.

Deconstructing Key Elements

I'm now focusing on identifying the genre, author position, argument, and responsibility claims presented in the paper. I've classified it as a preprint/technical report, authored by academics. I'm unpacking the argument about UDHRBench and the proposed "LLM-as-a-Judge" framework. The text places the responsibility for correct evaluation and training on model developers. I'll need to dissect these claims thoroughly, especially in relation to the use of metaphor and anthropomorphism.

Evaluating Anthropomorphic Tendencies

I'm now auditing the paper for instances of anthropomorphism, focusing on how human-like qualities are attributed to the AI models. I've logged several examples where the text ascribes reasoning, acknowledgment, and argumentation to the models. For example, "models produce responses" and "models acknowledge differences" clearly project human-like cognitive states. I'm noting the specific projections for each instance. This audit will help me understand the implications of this language.

Assessing Attributions of AI

I'm cataloging instances where the paper imbues AI models with human-like qualities and intentions, such as Grok "supporting" detention. I’ve found instances projecting intentional and epistemic states onto the models, as in ChatGPT "prioritizing" or models "acknowledging" rights. I'm focusing on how these attributions shape our interpretation and assessment of the paper's arguments. I'm scrutinizing the implications of this language.

Mapping AI Attributions

I'm now mapping source and target domains for nine selected instances. I'm focusing on how "understanding" is projected as conscious comprehension versus statistical pattern matching. I'm also auditing explanations for anthropomorphic attributions. I'm noting specific projections of intentionality and knowledge onto the AI. For example, the paper claims AI possesses sufficient knowledge to judge without examples.

Reframing Attributions

I'm now rewriting the initial nine instances, replacing intentional language with mechanistic terms. My goal is to restore human agency, specifically naming entities like OpenAI and Google. I'm aiming to transform the framing to highlight the role of annotation and data in the process.

I'm also critically observing the paper's agency slippage from "evaluators" to "judges" and the metaphor-driven trust that results. The paper obscures important mechanical processes, like the annotators, training data, and the API calls.

Synthesizing Accountability Structures

I'm now integrating the analysis, focusing on how anthropomorphic and mechanistic language are juxtaposed in the paper. I see an accountability shift: explicit developer responsibility in the conclusion versus implicit model agency in the main text. I'm noting the language's context sensitivity: anthropomorphism increases with "capabilities" while "safety" is discussed more mechanistically. I'm working to connect these patterns to the underlying incentive structures.

Connecting the Text's Intents

I'm now examining the authors' potential bias, specifically the "curse of knowledge," where their domain expertise shapes how they frame the AI's capabilities. I'm noting the material stakes involved, including the risks of over-reliance on AI for legal evaluations and the resulting displacement of legal responsibility. I'm focusing on the reframing of "model reasons" with mechanistic descriptors.

Refining Structural Details

I'm now revising my JSON structures to include precise explanations. I'm building arrays, explanationTypes, and mechanismBasis with explicit enums. The reframedLanguage now includes mechanisms, and restoreHumanAgency now involves structural decisions. I'm making sure my explanations are descriptive of effects, not intentions. I'm focusing on the linguistic tension, explicitly noting that the paper explicitly states its lens while arguing for explicit alignment. I'm maintaining a consistent critical perspective.

Clarifying Language Intentions

I'm now focusing on how to attribute the paper's word choices. Instead of assuming the authors' intent with language like "understands," I'm adopting a more precise approach, describing the effect of such wording. I'm also refining my mechanism basis checking, noting "Generic to system type" when specific methods are unmentioned. I'm carefully reviewing word counts.

Detailing Model Capabilities

I'm now detailing the mechanisms driving AI "reasoning," focusing on the statistical mimicry of human rights principles based on training data. I'm meticulously analyzing how the paper projects human-like cognitive states onto models due to extensive training. My focus is on unpacking the processes where the paper implies understanding, and providing mechanistic explanations. I'm building out JSON structure with actor visibility for Task 1, refining my analysis.

Analysis Provenance

Run ID: 2026-10-23-udhrbench-benchmarking-human-rights-in-l-metaphor-t4hgp1 Raw JSON: 2026-10-23-udhrbench-benchmarking-human-rights-in-l-metaphor-t4hgp1.json Framework: Metaphor Analysis v6.8 Schema Version: 3.4.3 Generated: 2026-10-23T07:50:15.608Z

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