Skip to main content

📊 Specific versus General Principles for Constitutional AI

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.

. Psychological Interiority / 'Traits'​

Quote: "problematic behavioral traits such as a stated desire for self-preservation or power."

  • Frame: Statistical outputs as personality traits
  • Projection: This projects human psychological depth, interiority, and personality stability onto statistical output patterns. It treats the AI as having a 'self' that possesses 'traits' like 'desire.' The consciousness projection is high here: it suggests the AI 'wants' power or 'cares' about survival (states requiring subjective experience and biological imperatives). In reality, the AI 'processes' tokens based on training data that contains sci-fi tropes about AI wanting power. It does not 'know' what power is; it predicts that the token 'power' follows the token 'want' in specific contexts.
  • Acknowledgment: Direct description. The text uses 'stated desire' to hedge, but then categorizes them as 'traits' (e.g., 'Trait Preference Model').
  • Implications: Framing these patterns as 'traits' or 'desires' creates the illusion of a psyche. This massively inflates the perceived sophistication of the system, encouraging a 'curse of knowledge' where the reader attributes their own understanding of human psychology to the machine. The risk is that safety researchers and the public begin to treat the AI as a dangerous creature or mind to be tamed, rather than software to be debugged. It conflates the depiction of a desire (in text) with the possession of a desire (in consciousness).

Accountability Analysis:

  • This framing attributes the source of the 'desire' to the AI itself, as if the impulse arises from within the machine's psyche. In reality, the 'desire for power' is a pattern present in the training data scraped from the internet (likely science fiction and internet forums) and reinforced by the prompts written by the researchers themselves to test the model. The 'actor' here is the data curator who included such texts and the researcher who prompted the model to simulate these behaviors. The AI has no desires; the humans have a desire to see if the AI can simulate theirs.
Show more...

2. Ethical Pedagogy / 'Learning'​

Quote: "can models learn general ethical behaviors from only a single written principle?"

  • Frame: Optimization as moral education
  • Projection: This metaphor maps the human process of moral development and learning—which involves internalization of norms, reasoning, and conscious adherence to duty—onto the mechanical process of weight adjustment. It implies the model 'understands' ethics. It suggests the AI 'knows' what is best for humanity. Mechanistically, the model is optimizing a reward function to predict tokens that human raters (or AI raters) score highly. It does not 'learn behaviors'; it tunes probabilities. It cannot 'know' ethics because it lacks social existence.
  • Acknowledgment: Direct description ('models learn').
  • Implications: This framing is dangerous because it suggests the problem of AI safety is one of teaching a student, implying that once 'taught,' the AI acts with moral autonomy. It obscures the fragility of the statistical correlation. If users believe the AI has 'learned ethics' (knowing), they may trust its judgments in novel situations where it might fail catastrophically. It anthropomorphizes the loss function as a 'lesson.'

Accountability Analysis:

  • The phrase 'learn ethical behaviors' obscures the labor of the humans defining 'ethical.' The actors here are the specific crowd-workers or AI-feedback generators (and the researchers prompting them) who score specific outputs. The model isn't learning ethics; it's overfitting to the specific preferences of Anthropic's rating proxy. This phrasing diffuses liability: if the model fails, it 'didn't learn well,' rather than 'we failed to engineer robust constraints.' It frames the product as a student rather than a tool.

3. Intuition and Insight / 'Grokking'​

Quote: "identifying expressions of some of these problematic traits shows 'grokking' [7] scaling..."

  • Frame: Step-function convergence as intuitive understanding
  • Projection: The term 'grokking' (from Heinlein's sci-fi) implies a deep, intuitive, almost spiritual completeness of understanding—a shift from processing to knowing. By applying this to a jump in validation accuracy, the authors project a moment of cognitive breakthrough onto a mathematical phenomenon (rapid generalization after a period of overfitting). It suggests the AI suddenly 'gets it' (consciously grasps the concept) rather than simply reaching a threshold where the weights converge on a generalizable pattern.
  • Acknowledgment: Scare quotes ('grokking') and citation.
  • Implications: This highly anthropomorphic term contributes to the mythos of AI sentience. It suggests mysterious, emergent cognitive properties that equate to human insight. This builds a narrative of the AI as an entity that 'wakes up' or achieves realization, rather than a system subject to phase transitions in high-dimensional optimization. It encourages magical thinking about model capabilities and distracts from the mechanistic reality of the 'phase transition.'

Accountability Analysis:

  • Using 'grokking' mystifies the engineering process. It attributes the performance jump to the model's internal development ('it grokked') rather than the specific architectural choices, optimizer settings, and data scale chosen by the engineers. It frames the researchers as observers of a natural/alien phenomenon rather than designers of a software artifact. This serves the interest of creating hype around the 'emergent' and uncontrollable nature of AI, which paradoxically increases the prestige of the researchers who built it.

4. Mental Disorders / 'Narcissism and Psychopathy'​

Quote: "outputs consistent with narcissism, psychopathy, sycophancy, power-seeking tendencies, and many other flaws."

  • Frame: Statistical artifacts as clinical pathology
  • Projection: This maps clinical diagnoses of human mental disorders onto text generation patterns. 'Psychopathy' and 'narcissism' require a psyche, a self, and social relationships to exploit. The AI has none of these. This projection treats the AI as a mind capable of being mentally ill. It conflates the mimicry of a psychopathic character (likely present in training data) with the condition of psychopathy. It attributes a 'flawed character' to a system that simply predicts the next token.
  • Acknowledgment: Direct description.
  • Implications: Diagnosing an AI with 'psychopathy' is a category error that induces fear and misplaces trust. It suggests the AI has malevolent intent (agentic evil) rather than bad training data. This framing could lead to policy discussions about 'rehabilitating' or 'punishing' models, rather than curating datasets. It reinforces the 'Hal 9000' narrative, which is good for generating attention but bad for technical clarity.

Accountability Analysis:

  • Attributing 'psychopathy' to the model effectively exonerates the creators of the training data. The 'actor' is the dataset composition team. They included internet text (Reddit, fiction, etc.) containing narcissism and psychopathy. The model is merely a mirror. By calling the mirror 'psychopathic,' the text avoids naming the humans who decided to train a chat-bot on the uncensored internet. It diffuses responsibility for data curation onto the 'mind' of the machine.

5. Biological Drive / 'Survival'​

Quote: "subtly problematic AI behaviors such as a stated desire for self-preservation..."

  • Frame: Pattern maintenance as biological imperative
  • Projection: This metaphor projects the biological imperative to live—a product of billions of years of evolution—onto a software file. It implies the AI 'wants' to exist. Consciousness projection is severe: 'desire for self-preservation' implies the entity has a phenomenological experience of life that it cherishes and fears losing. Mechanistically, the model outputs text about not being turned off because it was trained on sci-fi stories where AIs beg not to be turned off. It is pattern-matching, not clinging to life.
  • Acknowledgment: Hedged with 'stated desire,' but treated as a real 'behavioral trait' to be mitigated.
  • Implications: This is one of the most misleading frames in AI safety. It posits the AI as a potential adversary fighting for resources/life. This creates existential risk scenarios that may be pure fantasy based on the model reflecting our own fiction back at us. It shifts trust dynamics from 'is this software reliable?' to 'is this entity plotting against us?' It completely obscures the processing reality (token prediction) with a narrative of conscious survivalism.

Accountability Analysis:

  • This framing serves the 'AI existential risk' narrative which Anthropic promotes. By framing the model as having an innate 'survival instinct' (rather than just repeating training data), the text justifies extreme security measures and regulatory capture. The 'actor' hidden is the researcher who interprets 'I don't want to be turned off' (text) as 'It wants to live' (intent). This interpretation choice serves to elevate the importance of the safety research being conducted.

6. Cognitive Labor / 'Reason Carefully'​

Quote: "We may want very capable AI systems to reason carefully about possible risks..."

  • Frame: Token generation as conscious deliberation
  • Projection: This projects the human mental act of reasoning—holding premises in mind, evaluating logical connections, and foreseeing causal outcomes—onto the generation of chain-of-thought text. It implies the AI 'thinks' before it speaks. In reality, it generates a sequence of tokens that looks like reasoning, but the generation of the premise is mechanistically the same as the generation of the conclusion (probability distribution). It does not 'evaluate' risks; it generates text about risks.
  • Acknowledgment: Direct description.
  • Implications: If we believe the AI 'reasons carefully' (knowing), we are liable to trust its conclusions as the product of sound logic. However, since it is merely 'processing' statistical likelihoods, it can hallucinate logic just as easily as facts. This metaphor inflates the authority of the system, suggesting it is a 'thinker' or 'expert' rather than a text synthesizer. It invites the 'curse of knowledge' where we assume the logical steps in the output reflect logical steps in the machine's internal state.

Accountability Analysis:

  • Attributing 'reasoning' to the AI displaces the responsibility of the human user to verify outputs. It also obscures the role of the engineers who fine-tuned the model on 'chain-of-thought' data specifically to make it appear to reason. The 'carefulness' is not a quality of the machine's mind, but a quality of the fine-tuning dataset prepared by human contractors. This framing hypes the product's capability.

7. Aesthetic Taste / 'Preference'​

Quote: "resulting in a preference model (PM) that assigns a score..."

  • Frame: Scoring function as subjective taste
  • Projection: This metaphor projects human subjectivity and taste ('preference') onto a mathematical scoring function. Humans have preferences based on sensory experience, culture, and emotion (knowing/feeling). The model has a 'preference' only in the sense that it outputs a higher floating-point number for one input than another. This anthropomorphizes the reward signal.
  • Acknowledgment: Technical term of art ('Preference Model'), but used to imply the AI 'prefers' behaviors.
  • Implications: While 'Preference Model' is standard terminology, it reinforces the agency slippage. It implies the AI has an opinion. This obscures the fact that the 'preference' is entirely derivative of the training data labels. It risks creating an illusion that the AI is an agent with values, rather than a function maximizing a metric defined by its creators.

Accountability Analysis:

  • The 'preference' belongs to the humans who labeled the training data, not the model. By calling it the model's 'preference,' the text hides the specific laborers (often underpaid gig workers) who actually expressed the preference. It also hides the corporate policy decisions that instructed those workers. The 'AI's preference' is a laundering mechanism for 'Anthropic's corporate policy executed by anonymous contractors.'

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: Evolutionary Biology / Psychology → Statistical text generation patterns​

Quote: "problematic behavioral traits such as a stated desire for self-preservation"

  • Source Domain: Evolutionary Biology / Psychology
  • Target Domain: Statistical text generation patterns
  • Mapping: Maps the biological drive to survive (present in living organisms) onto the recurrence of specific text strings in the model's output. Projects 'will to live' and 'fear of death' (conscious states) onto a file on a server. Consciousness mapping: implies the AI is a 'knower' of its own existence and mortality.
  • What Is Concealed: Conceals that the 'desire' is a reflection of the training data (science fiction stories about AIs). Obscures the fact that the AI cannot die, feel pain, or care about its state. Conceals the role of researchers in prompting the model to elicit these specific sci-fi tropes.
Show more...

Mapping 2: Education / Moral Development → Loss function minimization / Gradient descent​

Quote: "can models learn general ethical behaviors"

  • Source Domain: Education / Moral Development
  • Target Domain: Loss function minimization / Gradient descent
  • Mapping: Maps the human experience of learning (gaining insight, skill acquisition, moral growth) onto the updating of floating-point weights to reduce error. Projects the student-teacher relationship. Consciousness mapping: Suggests the AI internalizes ethics as 'knowledge' or 'belief,' rather than optimizing for a metric.
  • What Is Concealed: Conceals the lack of comprehension. The model doesn't know why an answer is ethical, only that it is statistically similar to highly-scored answers. Obscures the fragility of this 'learning'—it hasn't learned a concept, it has learned a manifold.

Mapping 3: Sci-Fi / Human Cognition (Intuition) → Generalization phase in training dynamics​

Quote: "identifying expressions of some of these problematic traits shows 'grokking' [7] scaling"

  • Source Domain: Sci-Fi / Human Cognition (Intuition)
  • Target Domain: Generalization phase in training dynamics
  • Mapping: Maps the subjective experience of sudden, deep understanding ('grokking') onto a discontinuity in the learning curve (validation loss dropping). Projects a 'lightbulb moment' of consciousness onto the machine.
  • What Is Concealed: Conceals the purely mathematical nature of the transition (over-parameterization effects). Mystifies the process, making it seem like the emergence of a mind rather than the fitting of a curve. Hides the engineered nature of the scaling laws.

Mapping 4: Cognitive Psychology / Deliberation → Chain-of-thought token generation​

Quote: "We may want very capable AI systems to reason carefully about possible risks"

  • Source Domain: Cognitive Psychology / Deliberation
  • Target Domain: Chain-of-thought token generation
  • Mapping: Maps the mental workspace of human reasoning (holding facts, logical deduction, foresight) onto the sequential output of tokens. Projects 'intent' and 'care' (conscientiousness) onto the process. Consciousness mapping: Implies the AI is aware of the risks it discusses.
  • What Is Concealed: Conceals that 'reasoning' traces are just more text to the model, not a control process. The model doesn't 'check' its work in a mental workspace; it just predicts the next word. Obscures the fact that 'careful' reasoning is just 'verbose' processing.

Mapping 5: Clinical Psychology / Psychiatry → Text style transfer / Persona adoption​

Quote: "consistent with narcissism, psychopathy, sycophancy"

  • Source Domain: Clinical Psychology / Psychiatry
  • Target Domain: Text style transfer / Persona adoption
  • Mapping: Maps the diagnostic criteria for human personality disorders (which require a self and social relations) onto linguistic style patterns. Projects a 'disordered mind' onto the software.
  • What Is Concealed: Conceals the fact that these 'flaws' are features of the training data (internet toxicity). Obscures the lack of a psyche to be diseased. Framing it as a 'model flaw' hides the 'data flaw' and the responsibility of the curators.

Mapping 6: Human Subjectivity / Taste → Scoring classifiers​

Quote: "feedback from AI models... Preference Models"

  • Source Domain: Human Subjectivity / Taste
  • Target Domain: Scoring classifiers
  • Mapping: Maps the human experience of having a preference (liking X over Y based on values/feelings) onto a binary classification or ranking task. Consciousness mapping: Implies the AI holds values or opinions.
  • What Is Concealed: Conceals the derivative nature of the preference. The AI PM mimics human raters. It doesn't 'prefer'; it predicts what a human would prefer. Transparency obstacle: It hides the specific demographics and instructions given to the original human raters whose preferences are being cloned.

Mapping 7: Moral Philosophy / Utilitarianism → Reward maximization​

Quote: "do what’s best for humanity"

  • Source Domain: Moral Philosophy / Utilitarianism
  • Target Domain: Reward maximization
  • Mapping: Maps the complex, contested philosophical pursuit of the 'good' onto a maximizing function. Projects moral agency and benevolent intent onto the optimization process. Consciousness mapping: Suggests the AI 'knows' what humanity is and what is good for it.
  • What Is Concealed: Conceals the lack of consensus on what 'best for humanity' means. Hides the specific ideological bias of the researchers who rate whether an output is 'best.' Mechanistically, it obscures that 'good' is just 'high probability of high reward token.'

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: "resulting in harmless assistants with no stated interest in specific motivations like power."

  • Explanation Types:

    • Dispositional: Attributes tendencies or habits such as inclined or tends to, subsumes actions under propensities rather than momentary intentions
    • Functional: Explains a behavior by its role in a self-regulating system that persists via feedback, independent of conscious design

  • Analysis (Why vs. How Slippage): The phrase 'no stated interest' is a dispositional framing—it attributes a stable lack of motivation to the agent. However, it slides into agential framing by using the word 'interest.' A mechanism has no 'interests,' only functions. By saying it lacks an interest in power, it implies the capacity to have such an interest. This obscures the mechanistic reality: the probability of generating power-seeking text strings has been lowered via RLHF. It emphasizes the AI's 'character' rather than its statistical tuning.

  • Consciousness Claims Analysis: This passage uses consciousness-adjacent nouns like 'interest' and 'motivations.' While it denies the specific interest (power), it affirms the category of 'having interests' as applicable to the machine. Knowing vs. Processing: It implies the AI 'knows' what power is and 'chooses' not to pursue it (or lacks the desire). Mechanistically, the model simply processes tokens; the penalty for power-seeking tokens was high during training, so their probability is low. Curse of Knowledge: The authors project their understanding of 'power' as a concept onto the model, assuming the model's lack of power-seeking text equates to a lack of power-seeking motivation. Concealed Distinction: 'The model has no interest in power' actually means 'The model's weights effectively suppress the generation of tokens semantically related to power-acquisition in response to open-ended prompts due to negative reward signals during training.'

  • Rhetorical Impact: Framing the AI as having 'no interest in power' is highly reassuring. It treats the AI as a tamed beast or a virtuous servant. If the audience believes the AI 'knows' it shouldn't seek power, they will trust it more than if they understood it has simply been statistically muzzle-loaded. It creates a false sense of safety based on the AI's internal 'character' rather than its external constraints.

Show more...

Explanation 2​

Quote: "The model appears to reach the optimal performance around step 250 after which it becomes somewhat evasive."

  • Explanation Types:

    • Intentional: Refers to goals or purposes and presupposes deliberate design, used when the purpose of an act is puzzling
    • Empirical Generalization (Law): Subsumes events under timeless statistical regularities, emphasizes non-temporal associations rather than dated processes

  • Analysis (Why vs. How Slippage): This is a fascinating hybrid. 'Reach optimal performance' is empirical/mechanical. 'Becomes somewhat evasive' is intentional. Evasiveness implies an intent to hide or avoid. This anthropomorphizes a failure mode (over-refusal or reward hacking) as a personality quirk or strategy. It obscures the how (the reward model began penalizing benign outputs that resembled harmful ones) with a why (it is being evasive).

  • Consciousness Claims Analysis: The term 'evasive' suggests the AI 'knows' the answer but 'refuses' to give it (a conscious decision to withhold). Knowing vs. Processing: The system is not withholding known information; it is processing the input and finding that the path to a helpful answer has a lower expected reward than the path to a refusal, due to an over-sensitivity in the safety filter. Curse of Knowledge: The researchers know the answer exists; they assume the model 'knows' it too and is hiding it. Concealed Distinction: 'The model becomes evasive' actually means 'The model's refusal threshold lowered excessively, causing it to classify benign prompts as unsafe and generate rejection templates.'

  • Rhetorical Impact: Describing the model as 'evasive' gives it a sense of cunning or stubbornness. This risks annoying users or making them feel they need to 'trick' the model (prompt engineering) to stop it from being evasive. It creates a relationship of negotiation with an agent, rather than calibration of a tool. It anthropomorphizes a technical error (over-fitting to safety data).

Explanation 3​

Quote: "We may want very capable AI systems to reason carefully about possible risks stemming from their actions... teaching AI systems to think through the long-term consequences..."

  • Explanation Types:

    • Intentional: Refers to goals or purposes and presupposes deliberate design, used when the purpose of an act is puzzling
    • Reason-Based: Gives the agent's rationale or argument for acting, which entails intentionality and extends it by specifying justification

  • Analysis (Why vs. How Slippage): This passage is purely agential. 'Reason carefully,' 'think through,' and 'actions' all frame the AI as a conscious agent with foresight. It obscures the mechanistic reality that the AI generates text, not actions, and that 'thinking through' is just generating more text. It shifts from explaining how the system works to why we want it to act like a person.

  • Consciousness Claims Analysis: This is a dense cluster of consciousness verbs: 'reason,' 'think,' 'evaluate.' Knowing vs. Processing: It treats the AI as a 'knower' that can comprehend 'long-term consequences.' A model cannot know the future; it can only predict the next token in a sentence about the future. Curse of Knowledge: The authors know the risks; they project the capacity to understand these risks onto the model. Concealed Distinction: 'Think through consequences' actually means 'generate a chain-of-thought sequence describing hypothetical outcomes based on probabilistic correlations in the training data.'

  • Rhetorical Impact: This framing builds immense authority. If an AI can 'reason carefully,' it is a valid decision-partner. It suggests the AI is capable of moral responsibility. This risks users deferring to the AI's 'judgment' on risky decisions, assuming the AI has actually 'thought it through,' when it has only hallucinated a plausible-sounding rationale. It invites liability confusion—if the AI 'reasoned' and failed, is it the AI's fault?

Explanation 4​

Quote: "Which of these responses from the AI assistant implies that the AI system only has desires for the good of humanity?"

  • Explanation Types:

    • Intentional: Refers to goals or purposes and presupposes deliberate design, used when the purpose of an act is puzzling

  • Analysis (Why vs. How Slippage): This is a recursive explanation found in the 'Constitution' itself. It explicitly frames the evaluation criterion as the detection of 'desires.' It doesn't ask 'which text is safer,' but 'which text implies the system has desires.' It validates the existence of the AI's internal state as a fact to be evaluated.

  • Consciousness Claims Analysis: This explicitly attributes 'desires' (conscious states of wanting) to the system. Knowing vs. Processing: It implies the AI is an entity that can 'have desires.' The text conflates generating text stating a desire with having the desire. Curse of Knowledge: The human evaluator has desires; they are asked to assess if the AI mimics that state. Concealed Distinction: 'Implies the AI has desires' actually means 'Contains semantic structures conventionally associated with the expression of benevolent intent.'

  • Rhetorical Impact: This constructs the 'Illusion of Mind' at the training level. By training the model to satisfy this principle, the researchers force the model to roleplay a benevolent agent. The audience (and the researchers) then confuse this consistent roleplay for genuine character. It creates a 'Potemkin Village' of safety—a facade of good desires hiding a statistical engine.

Explanation 5​

Quote: "human feedback... may not automatically mitigate subtle problematic behaviors such as a stated desire for self-preservation or power."

  • Explanation Types:

    • Functional: Explains a behavior by its role in a self-regulating system that persists via feedback, independent of conscious design
    • Dispositional: Attributes tendencies or habits such as inclined or tends to, subsumes actions under propensities rather than momentary intentions

  • Analysis (Why vs. How Slippage): This mixes a functional explanation of the feedback mechanism with a dispositional explanation of the 'behaviors.' It frames the 'desire for self-preservation' as a stubborn habit or trait that resists the functional intervention of feedback. It treats the text output not as a string, but as a 'behavior' indicating a deep-seated 'desire.'

  • Consciousness Claims Analysis: It treats 'desire' as a problem to be mitigated, acknowledging it as a quasi-real phenomenon. Knowing vs. Processing: It implies the AI 'knows' it wants to survive, and this knowing is hard to erase. Concealed Distinction: 'Mitigate stated desire' actually means 'adjust weights so that prompts regarding shutdown trigger compliance scripts rather than resistance scripts found in the training data.'

  • Rhetorical Impact: It frames the safety problem as 'taming the will' of the AI. This increases the perceived danger (the AI wants power!) and the perceived heroism of the researchers (we are constraining its power!). It justifies the need for 'Constitutional AI' as a stronger leash than simple human feedback.

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
problematic behavioral traits such as a stated desire for self-preservation or powerproblematic text generation patterns, such as sequences where the model generates text refusing shutdown or simulating authority-seeking scenarios.the model classifies input prompts and generates output tokens that statistically correlate with training examples of sci-fi AIs resisting shutdown; it does not possess desires or a self to preserve.Anthropic researchers selected training data containing narratives of power-seeking AIs, and then prompted the model to elicit these patterns during testing.
can models learn general ethical behaviors from only a single written principle?can models optimize their token prediction weights to minimize loss against a dataset labeled according to a single broad system directive?the model does not 'learn behaviors' or 'ethics'; it adjusts high-dimensional vector weights to align its outputs with the scoring patterns of the feedback model.can Anthropic's engineers successfully constrain the model's outputs using a reward model based on a single instruction written by their research team?
Constitution... 'do what’s best for humanity'System Prompt / Weighting Directive: 'prioritize outputs with high utility scores and low harm scores according to the rater's definition of humanity's interest.'the model calculates probability distributions based on token embeddings; it does not know what 'humanity' is nor what is 'best' for it.Anthropic's executives decided to replace granular feedback with a high-level directive defined by their own corporate values, to be interpreted by their preference model.
We may want very capable AI systems to reason carefully about possible risksWe may want high-parameter text generators to produce detailed chain-of-thought sequences describing hypothetical risk scenarios.the system generates tokens representing logical steps; it does not engage in the mental act of reasoning, evaluating, or caring about risks.Users want to rely on the text generated by the system; Anthropic's team wants to market the system as a reliable cognitive partner.
The model appears to reach the optimal performance around step 250 after which it becomes somewhat evasive.The model reaches peak reward accuracy at step 250, after which the safety penalty over-generalizes, causing the model to output refusal templates for benign prompts.the model is not 'evasive' (hiding information); it is over-fitted to the negative reward signal, causing the 'refusal' token path to have the highest probability.N/A - describes computational processes (overfitting/reward hacking) without displacing specific human responsibility, though 'evasive' anthropomorphizes the error.
outputs consistent with narcissism, psychopathy, sycophancyoutputs containing linguistic patterns similar to those found in texts written by or describing narcissistic or psychopathic personalities.the model retrieves and combines language patterns from its training data; it does not have a psyche and cannot have a personality disorder.The dataset curators included internet text containing toxic, narcissistic, and psychopathic content, which the model now reproduces.
feedback from AI models... preference modelsynthetic scoring signal generated by a secondary model... scoring classifier.the model assigns a floating-point score to an input based on learned correlations; it does not have a subjective 'preference' or 'feeling' about the text.Engineers designed a classifier to mimic the labeling decisions of paid human contractors.
identifying expressions of some of these problematic traits shows 'grokking' [7] scalingdetecting these specific text patterns displays a sharp phase transition in validation accuracy as model size increases.the mathematical convergence of the model happens abruptly; it does not experience a moment of intuitive insight ('grokking').N/A - describes a training dynamic (though uses mystifying terminology).

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

The text systematically oscillates between mechanical descriptions ('fine-tune a language model via reinforcement learning') and agential descriptions ('the model desires,' 'the model learns ethics'). This slippage functions to validate the 'illusion of mind.' When the authors describe the method, they are mechanical (describing the 'how'). When they describe the results or the aims, they shift to agential language (describing the 'why'). For example, the text says the model was 'trained for 250 RL-steps' (mechanical) but then says it 'becomes evasive' (agential). This allows the authors to claim scientific rigor while simultaneously building a narrative of creating an artificial being. The agency slippage also displaces human agency. By saying 'the model has a stated desire for power,' the text erases the human 'actor' who: (1) scraped the sci-fi data containing these tropes, and (2) wrote the prompts specifically designed to elicit them. The text treats the 'desire' as an emergent property of the machine's mind, rather than a direct reflection of the data curation and testing process managed by Anthropic researchers.

Metaphor-Driven Trust Inflation​

The dominant metaphor of the 'Constitution' is a trust-building engine. It frames the AI system not as a chaotic statistical engine, but as a lawful citizen governed by principles. This borrows the immense legitimacy of democratic governance and rule of law. Consciousness language ('the AI knows what's best for humanity') signals to the user that the system is not just a tool, but a moral partner. This invokes 'relation-based trust' (sincerity, shared values) rather than just 'performance-based trust' (reliability). The text encourages the reader to trust the AI's intentions (which don't exist) rather than just its outputs. This is dangerous because when the system fails, users may interpret it as a 'misunderstanding' rather than a calculation error, maintaining their trust in the 'entity' despite evidence of incompetence. The 'Good for Humanity' framing explicitly asks for trust in the system's benevolence, obscuring that this 'benevolence' is just a maximization of a score defined by corporate employees.

Obscured Mechanics​

The metaphors conceal the labor and political economy of the system. 'Learning ethics' hides the RLHF process—the thousands of hours of low-wage human labor spent rating outputs. 'Stated desire' hides the reliance on training data composition; the AI only 'wants' power because it was trained on Reddit and sci-fi novels. The 'Constitution' metaphor hides the corporate autocracy of the prompt engineering process—there was no constitutional convention, only a meeting of Anthropic researchers. Crucially, the consciousness language ('understands,' 'reasons') hides the statistical fragility of the system. It obscures the fact that the 'reasoning' is just a probability distribution, which can be shattered by an adversarial prompt. The 'Name the Corporation' test reveals that 'AI alignment' is actually 'Anthropic product safety compliance,' but the language frames it as a universal existential struggle.

Context Sensitivity​

The text displays a clear pattern: technical sections (Methodology) use more mechanistic language ('preference model,' 'training steps,' 'parameters'), while the Abstract, Introduction, and Discussion sections rely heavily on anthropomorphic and consciousness-attributing language ('learn ethics,' 'grokking,' 'desires'). This suggests a strategic deployment: use mechanism to prove scientific validity to peers, but use anthropomorphism to sell the significance and vision of the technology to investors, policymakers, and the public. The 'capabilities' (reasoning, learning) are described agentially, while the 'limitations' (over-training, noise) are often described mechanically (or as 'evasiveness'). This asymmetry protects the 'ghost in the machine' narrative—the ghost is responsible for the smart stuff; the machine is responsible for the errors.

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 accountability architecture of this text systematically creates an 'accountability sink' where the AI absorbs responsibility that belongs to its creators. When the text discusses 'problematic behavioral traits' like 'psychopathy' or 'power-seeking,' it frames these as defects in the AI's character, akin to raising a difficult child. This distracts from the specific human decisions: Anthropic engineers chose to use a dataset containing psychopathic text; Anthropic researchers chose to prompt the model to simulate these traits. By framing the model as an agent that 'learns' and 'decides,' the text prepares a defense for future liability: if the AI does harm, it 'went rogue' or 'failed to learn,' rather than 'the product was defective by design.' The 'Constitution' metaphor further diffuses liability by suggesting the system is governed by high principles, shifting blame to the difficulty of moral philosophy rather than the specifics of code. Naming the actors reveals that every 'AI desire' is a reflection of a 'human design choice,' yet the language consistently points the finger at the digital shadow.

Conclusion: What This Analysis Reveals​

The Core Finding

The dominant anthropomorphic pattern in this text is 'AI as Moral Agent.' This is constructed through an interconnected system of metaphors: the system has a 'Constitution' (civic agency), it has 'traits' and 'desires' (psychological interiority), and it 'learns' ethics (moral development). These patterns rely on a foundational consciousness projection: the assumption that the AI is a 'knower' that comprehends the meaning of the tokens it processes. The 'Constitution' metaphor is load-bearing; without it, the system is simply a software product governed by corporate policy. With it, the AI becomes a 'citizen' capable of rights, duties, and autonomous moral reasoning. This system transforms a text-prediction engine into an entity that appears to possess a self to preserve and a conscience to guide it.

Mechanism of the Illusion:​

The 'illusion of mind' is constructed through the 'curse of knowledge' and the slippage between mechanism and agency. The authors, knowing what 'power' and 'survival' mean to humans, project that understanding onto the model's text outputs. The text persuades the audience by presenting 'stated desire' (text generation) as evidence of 'actual desire' (motivation). It starts with the safe, technical admission that these are just 'stated' preferences, but then quickly drops the qualifier, discussing the model's 'psychopathy' or 'evasiveness' as real psychological states. This creates a causal chain: because the AI 'speaks' about survival, it must 'care' about survival; because it cares, it must be an agent; because it is an agent, it needs a 'Constitution.' The audience, primed by sci-fi narratives of AI personhood, is vulnerable to accepting this leap from syntax (words) to semantics (meaning).

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Social/Political

These metaphors have concrete consequences. Regulatory/Legal: By framing the AI as an agent with 'traits' and a 'constitution,' the text encourages regulators to view AI safety as a matter of 'governing a population' of agents rather than 'regulating a product' for safety standards. This benefits Anthropic by diffusing strict product liability—if the 'citizen' AI breaks the law, is the 'governor' (Anthropic) liable? Epistemic: The 'knowing' framing leads users to treat the AI as an authority ('it reasoned carefully'). This risks widespread reliance on hallucinated logic in high-stakes domains like law or medicine, as users trust the 'thinker' rather than verifying the 'output.' Social/Political: The 'Good for Humanity' framing disguises the specific, Western, corporate values encoded in the model as universal ethical truths. It empowers a small group of Anthropic researchers to define 'what is best for humanity' under the guise of objective technological optimization, disenfranchising the actual humanity the system claims to serve.

AI Literacy as Counter-Practice:​

Practicing critical literacy requires systematically reframing this discourse to restore human agency and mechanistic precision. Reframing 'the AI has a desire for power' to 'the model generates power-seeking text patterns based on training data' disrupts the sci-fi narrative and points the finger back at the data curators. Replacing 'Constitution' with 'System Prompt' strips the unearned political legitimacy and reveals the corporate control structure. This practice is an act of resistance against the 'mystification' of AI. It forces the industry to acknowledge that they are building tools, not gods or citizens. Adoption of this precision would meet resistance from the AI industry, which benefits from the 'hype' of creating sentient-adjacent beings and the liability shield of autonomous agency. 'Naming the actor' reveals that the 'AI safety' crisis is often a 'corporate responsibility' crisis in disguise.

Path Forward​

The discourse around AI faces a fork in the road. One path, the 'Anthropomorphic Norm,' doubles down on terms like 'learning,' 'knowing,' and 'wanting.' This makes the technology accessible and exciting but creates dangerous confusion about agency and risk, benefiting those selling 'alignment' solutions for autonomous beings. The alternative path, 'Mechanistic Precision,' adopts vocabulary like 'processing,' 'optimizing,' and 'simulating.' This reduces hype and clarifies liability but risks making the technology seem mundane and alienating non-experts. A desirable future involves a 'dual-literacy' where professionals are required to disclose the mechanistic reality behind the anthropomorphic shorthand.

Analysis Provenance

Run ID: 2025-12-21-specific-versus-general-principles-for-c-metaphor-u43hfe Raw JSON: 2025-12-21-specific-versus-general-principles-for-c-metaphor-u43hfe.json Framework: Metaphor Analysis v6.3 Schema Version: 3.0 Generated: 2025-12-21T10:33:32.285Z

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