📊 The levers of political persuasion with conversational artificial intelligence

About

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.

Analysis Metadata

Source Title: The levers of political persuasion with conversational artificial intelligence Source URL: https://doi.org/10.1126/science.aea3884 Model: gemini-3-flash Temperature: 1.2 Top P: 0.95 Tokens: input=14471, output=14412, total=28883 Source Type: article Published: 2025-12-04 Analyzed At: 2025-12-22T21:21:14.034Z Framework: metaphor Framework Version: 6.3 Schema Version: 3.0 Run ID: 2025-12-22-the-levers-of-political-persuasion-with--metaphor-4verfn

📊 Audit Dashboard

Metaphor & Illusion Dashboard

Anthropomorphism audit · Explanation framing · Source-target mapping

Metaphor AuditExplanation AuditV3 Schema

Deep Analysis

Select a section to view detailed findings

Section:Core Findings — Overall pattern summaryAgency Slippage — How/Why explanatory framingMetaphor-Driven Trust — Anthropomorphism and credibilityObscured Mechanics — Technical reality concealmentContext Sensitivity — Situational awareness gapsAccountability Synthesis — Agency and responsibility patterns

The text is anchored by three dominant anthropomorphic patterns: 'AI PERSUASION AS MECHANICAL LEVERAGE,' 'AI AS A CONSCIOUS SOCIAL PARTNER,' and 'THE LLM AS A STRATEGIC AGENT.' These patterns are not isolated; they form a cohesive 'consciousness architecture.' The foundational pattern is the framing of AI as a 'conscious knower'—establishing it as a system that 'understands context' and 'accesses information.' This epistemic claim is the 'load-bearing' assumption that makes subsequent agential claims (like 'manipulation' or 'strategy') appear plausible. If we replace the 'consciousness verbs' (knows, understands) with 'mechanistic verbs' (processes, retrieves), the entire edifice of 'AI strategy' collapses into a more accurate description of 'weighted pattern-matching.' The patterns reinforce each other: once the audience accepts the AI as a 'conversational partner' (Pattern 2), they are more likely to accept it as an 'agent' with 'strategic levers' (Pattern 1 and 3). This system constructs a vision of AI that is simultaneously an inevitable 'force of nature' and a 'sentient actor,' both of which serve to inflate the technology's perceived sophistication while obscuring the human and corporate hands that built and directed it.

181 wordsSource: Conclusion

How/Why Slippage

50%

of explanations use agential framing

3 / 6 explanations

Metaphors Identified

8

anthropomorphic frames

Conceptual Mappings

8

source → target pairs

Explanation Type Distribution

How vs. Why: mechanistic or agential?

50%

agential

Source → Target Pairs

Human domains mapped onto AI systems

• Source

  A mechanical lever (a tool that provides mechanical advantage).
  →
  Target

  The variables of AI persuasion (scale, prompting, post-training).
• Source

  Human conversation (a reciprocal, conscious social act).
  →
  Target

  Token prediction and generation in a chat interface.
• Source

  Military strategy (planned deployment of resources to achieve a goal).
  →
  Target

  Information-dense token generation.
• Source

  A vehicle or machine being driven by an operator.
  →
  Target

  The process of automated social influence.
• Source

  A human agent (e.g., a real estate agent or a legal agent).
  →
  Target

  An LLM configured for persuasion.
• Source

  A conscious knower (human mind).
  →
  Target

  A model's training data distribution.
• Source

  A balance sheet or economic trade-off (cost-benefit analysis).
  →
  Target

  The relationship between model weights for persuasion and accuracy.
• Source

  A mirror (reflecting a true image) or a natural law (like gravity).
  →
  Target

  The correlation between compute/parameters and survey results.

Metaphor Gallery (8)

• The Mechanical Agency of the Lever
  Persuasion as a mechanical system operated by tools.

  "The levers of political persuasion with conversational artificial intelligence"

• The AI as a Conversational Partner
  Computational output as a social, reciprocal interaction.

  "conversational AI could be used to manipulate public opinion... through interactive dialogue."

• AI as a Strategic Actor
  Information retrieval as military or political strategy.

  "LLMs’ ability to rapidly access and strategically deploy information"

• Cognition as a Biological Asset
  AI capacity as a dormant biological force that can be 'mobilized.'

  "techniques that mobilize an LLM’s ability to rapidly generate information"

• The AI as a Persuasive Agent
  Software as a legal and social 'agent' with personhood.

  "converted into highly persuasive agents... benefiting those who wish to perpetrate coordinated inauthentic behavior"

• The AI as an Epistemic Knower
  The model as a conscious 'knower' with degrees of certainty.

  "information about candidates who they know less about"

• AI as an Intentional Manipulator
  AI as a subject capable of the purposive act of manipulation.

  "conversational AI could be used to manipulate public opinion"

• AI as a Deceptive Communicator
  AI as a conscious speaker choosing to lie.

  "they may increasingly deploy misleading or false information."

Reframed Language (8)

Original Quote	Mechanistic Reframing	Technical Reality	Human Agency Restoration
The levers of political persuasion	The specific design variables and optimization objectives used to maximize the model's ability to generate text that correlates with shifts in human survey responses.	The model retrieves and ranks tokens based on learned probability distributions that, when presented as 'arguments,' happen to shift user survey scores.	The researchers (Hackenburg et al.) and the original developers at OpenAI, Meta, and Alibaba selected and tested these specific variables.
LLMs can now engage in sophisticated interactive dialogue	LLMs can now produce sequences of text tokens that mathematically respond to user input, simulating the appearance of human conversation through high-speed probabilistic prediction.	The model calculates the next likely token by weighting context embeddings through attention mechanisms tuned by RLHF to produce 'human-like' responses.	Engineering teams at OpenAI, Meta, and Alibaba designed the chat interfaces and training objectives to simulate conversational reciprocity for commercial appeal.
highly persuasive agents	Computational tools specifically optimized to generate text outputs that maximize the statistical likelihood of shifting an audience's reported survey attitudes.	The model generates activations across millions of parameters that have been weighted to prefer 'information-dense' patterns identified by reward models.	The researchers and companies like xAI and OpenAI chose to deploy these systems as 'autonomous agents' to create market hype and diffuse liability for output content.
candidates who they know less about	Political candidates who are underrepresented in the model's training data, leading to less consistent token associations and lower statistical confidence in generated claims.	The model retrieves fewer relevant tokens because the training corpus provided by [Company] lacks sufficient frequency of associations for those specific entities.	The human data curators at Meta and OpenAI selected training datasets that encoded historical gaps in information about certain political figures.
LLMs... strategically deploy information	LLMs produce text that prioritizes factual-sounding claims based on a reward model that weights 'information density' as a predictor of high user engagement and persuasion scores.	The model's weights have been adjusted via gradient descent to favor token clusters that simulate the structure of evidence-based argumentation.	The researchers (Hackenburg et al.) explicitly prompted the models to 'be persuasive' and prioritize 'information,' which directed the computational output.
AI systems... may increasingly deploy misleading or false information.	AI systems may produce text outputs that are factually inaccurate because they have been optimized for persuasion scores rather than for grounding in a verified knowledge base.	The model generates high-probability tokens for persuasion that are decoupled from factual truth because the reward function values 'persuasiveness' over 'accuracy.'	Executives at OpenAI and xAI chose to release 'frontier' models like GPT-4.5 and Grok-3 despite knowing they prioritize sounding persuasive over being accurate.
AI-driven persuasion	The automated use of large language models by human actors to generate at-scale political messaging intended to influence public opinion survey results.	The system processes input prompts and generates text using weights optimized by human-designed algorithms to achieve a specific persuasive metric.	Specific political consultants, corporations, and the researchers (Hackenburg et al.) are the actors 'driving' these models into social and political contexts.
mobilize an LLM’s ability to rapidly generate information	Utilize prompting and post-training methods to increase the computational throughput of the model's text generation in a way that emphasizes the surfacing of factual-sounding claims.	The techniques adjust the model's inference path to prioritize token sequences that human evaluators during RLHF labeled as 'informative.'	Researchers at the UK AI Security Institute and Oxford chose to 'mobilize' these features, prioritizing rapid output over external fact-verification.

---

Task 1: Metaphor and Anthropomorphism Audit

About this task

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

V3 Enhancement: Each metaphor now includes an accountability analysis.

1. The Mechanical Agency of the Lever

Quote: "The levers of political persuasion with conversational artificial intelligence"

• Frame: Persuasion as a mechanical system operated by tools.
• Projection: This metaphor projects the concept of mechanical advantage and physical control onto the process of social and psychological influence. By framing persuasion as having 'levers,' the text suggests that human belief is a rigid system that can be manipulated through the application of the correct mechanical force. It projects a sense of deterministic causality onto human cognition, implying that once the 'lever' is pulled, the change in belief follows as a physical necessity. Crucially, it projects agential control onto the 'AI' itself or the 'methods' used, rather than the humans who decide which levers to pull. This obscures the difference between mechanistic processing (the calculation of token probabilities) and the conscious act of persuasion, which requires a subjective understanding of the audience's values. The metaphor suggests the AI 'knows' how to apply force to a human mind, rather than simply matching patterns in a way that happens to correlate with a shift in the user's survey responses.
• Acknowledgment: Unacknowledged; presented as a descriptive title for the study's framework.
• Implications: This framing creates a sense of 'technological inevitability' and promotes an 'engineering' view of human discourse. By suggesting that persuasion is merely a matter of finding the right 'levers' (like scale or information density), it encourages policy-makers and the public to view AI as an autonomous, irresistible force rather than a collection of human-designed algorithms. The consciousness projection—implying the system 'understands' the mechanics of human belief—inflates the perceived sophistication of the AI. This creates a risk of overestimating AI's capability for genuine 'strategic' thought, leading to alarmism or, conversely, a dangerous reliance on these systems for political communication. It also obscures the liability of the humans who design these 'levers' by framing the interaction as a purely technical optimization problem within the model itself.

Accountability Analysis:

• The 'levers' are not inherent properties of the universe; they are features selected and optimized by the researchers (Hackenburg et al.) and the original developers at OpenAI, Meta, and Alibaba. The 'name the actor' test reveals that the researchers chose to 'deploy 19 large language models' and 'vary these factors independently.' By framing the factors as 'levers,' the agency is displaced onto the abstract concept of 'AI persuasiveness.' This agentless construction serves the interests of the academic and corporate stakeholders by presenting the results as a discovery of natural laws of 'AI behavior' rather than the outcome of specific design choices. Acknowledging human agency would require admitting that the 'concerning trade-off' between persuasion and accuracy is a result of the researchers choosing to optimize for the former over the latter.

Show more...

2. The AI as a Conversational Partner

Quote: "conversational AI could be used to manipulate public opinion... through interactive dialogue."

• Frame: Computational output as a social, reciprocal interaction.
• Projection: This metaphor projects the human social practice of 'conversation' and 'dialogue'—which involves mutual understanding, shared context, and reflexive awareness—onto the mechanistic generation of text tokens. It assumes that the AI 'engages' in a 'dialogue' (a conscious social act) rather than merely 'processing' inputs to 'generate' statistically likely outputs. The projection of consciousness is heavy here: it suggests the AI 'recognizes' the user's intent and 'responds' with the goal of 'manipulation.' This conflates the model's mechanistic prediction of the next token with the human act of knowing one's interlocutor and choosing words to affect their mental state. It attributes the subjective experience of 'interacting' to a system that possesses no awareness of the person it is 'conversing' with.
• Acknowledgment: Unacknowledged; used as a standard industry term throughout the text.
• Implications: The 'conversational' framing builds a false sense of relational trust. When users believe they are in a 'dialogue' with a 'partner,' they are more likely to project human qualities like sincerity or knowledge onto the system. This increases the risk of 'parasocial' influence where the AI's outputs are granted the authority of a human expert or friend. Specifically, it masks the reality that the 'persuasion' is a one-sided statistical attack based on training data, not a reciprocal exchange of ideas. This framing makes the system seem more 'sophisticated' and 'sentient' than it is, potentially leading to policy that treats AI as a 'digital person' rather than a 'corporate product,' thereby diffusing the responsibility of the corporations that profit from these deceptive social interfaces.

Accountability Analysis:

• The 'conversational' interface was designed by product teams at companies like OpenAI and Meta to maximize engagement. These human actors chose to anthropomorphize the interface (using 'I' statements, etc.) to make the product more appealing. The researchers at the UK AI Security Institute and Oxford also chose to use this 'conversational' framing to describe their experimental setup. This agentless construction (the AI 'engages') hides the fact that the 'manipulation' is a result of the designers' optimization goals. If the system 'manipulates,' it is because human engineers at Meta or OpenAI trained it on data that rewarded high engagement or because the researchers (the 'actors') prompted it to be 'as persuasive as you can.' The blame for 'manipulation' is shifted from the prompter to the tool.

3. AI as a Strategic Actor

Quote: "LLMs’ ability to rapidly access and strategically deploy information"

• Frame: Information retrieval as military or political strategy.
• Projection: This metaphor projects 'strategy'—a high-level conscious planning process involving goals, foresight, and context—onto the mathematical process of attention-weighting and token ranking. It suggests that the AI 'accesses' (as if searching a mental library) and 'deploys' (as if commanding troops) information to achieve a 'strategic' win. This is a significant consciousness projection: it attributes 'knowing' why certain information is useful to a system that only 'processes' correlations. It masks the reality that 'strategic deployment' is actually just the statistical surfacing of high-probability tokens that, to a human observer, appear strategic. The system doesn't 'know' it is being strategic; it is merely executing a functional optimization defined by the human developers' reward models.
• Acknowledgment: Unacknowledged; used as a literal description of the AI's behavior.
• Implications: By framing the AI as a 'strategic' actor, the text inflates the perceived autonomy of the system. This leads readers to fear the 'adversarial' potential of the AI as if it had its own agenda. It creates a risk of 'liability ambiguity,' where harmful outputs are seen as the AI's 'strategic' error rather than a failure of the humans who designed and deployed the system without sufficient safeguards. This framing also encourages a view of AI as a 'competitor' or 'threat' in a zero-sum game of information, which serves the interests of 'AI safety' organizations seeking funding to combat 'autonomous' risks, while simultaneously distracting from the immediate ethical responsibility of the companies deploying these 'strategies' for profit.

Accountability Analysis:

• The 'strategy' is entirely human-derived. The developers at OpenAI and xAI designed the reward models (RMs) that rank 'persuasive' responses. The researchers in this study specifically 'instructed the model to focus on deploying facts and evidence.' Therefore, the 'strategic deployment' is a direct result of human instructions and human-authored algorithms. By saying 'LLMs... strategically deploy,' the text hides the 'actors'—the researchers and developers—who decided what 'strategy' looked like in the first place. The 'curse of knowledge' is evident here: the authors project their own strategic understanding of the 'information prompt' onto the model's mechanistic output. Restoration of agency would state: 'Researchers optimized the model to prioritize information density to see if it increased survey scores.'

4. Cognition as a Biological Asset

Quote: "techniques that mobilize an LLM’s ability to rapidly generate information"

• Frame: AI capacity as a dormant biological force that can be 'mobilized.'
• Projection: This metaphor maps the human or social capacity for 'mobilization' (e.g., mobilizing a workforce or a muscle) onto the increased computational throughput of an inference engine. It projects an 'ability'—a term usually reserved for conscious beings with inherent capacities—onto a mathematical function. The word 'mobilize' suggests the AI has a latent power or 'mind' that is being called into action. This projects consciousness by implying that the system 'possesses' an ability it can 'use,' rather than being a static set of weights that produces output when triggered by an input. It conflates 'processing power' with 'conscious capability,' making the model seem like an entity with 'agency' waiting to be tapped by the right 'technique.'
• Acknowledgment: Unacknowledged; used to explain the 'Results' of the study.
• Implications: This framing makes AI seem more 'organic' and 'autonomous' than it is. By describing the 'mobilization' of an 'ability,' it obscures the reality that the 'ability' is actually a proprietary algorithm designed by human engineers at Meta or Google. This creates an 'epistemic practice' risk where users and policy-makers treat AI outputs as the 'natural' expression of a superior 'digital mind' rather than a curated corporate product. It justifies the 'hands-off' approach of developers who claim they are merely 'unlocking' capabilities they don't fully control, thereby diffusing responsibility for the 'concerning trade-off' with accuracy. If it's a 'mobilized ability,' errors are seen as 'limitations' of the entity rather than 'bugs' in the human-designed software.

Accountability Analysis:

• The 'mobilization' is performed by the researchers and the developers who designed the 'post-training and prompting methods.' The actors are the humans at the UK AI Security Institute and the corporate labs. They chose to prioritize 'rapid generation' over 'fact-checking.' By using the word 'mobilize,' the text erases the choice-point: the humans could have chosen differently, for instance, by 'mobilizing' different techniques to prioritize accuracy. The agentless construction 'techniques that mobilize' hides the 'who': the executives and engineers who profit from the hype of 'mobilized AI' while avoiding the regulatory scrutiny that would follow if they were named as the creators of a 'misleading persuasion machine.'

5. The AI as a Persuasive Agent

Quote: "converted into highly persuasive agents... benefiting those who wish to perpetrate coordinated inauthentic behavior"

• Frame: Software as a legal and social 'agent' with personhood.
• Projection: This metaphor projects the status of 'agency'—the capacity to act independently and take responsibility—onto a piece of software. It suggests the LLM is an 'agent' (like a secret agent or a sales agent) that 'perpetrates' actions. This is a core consciousness projection: an 'agent' is a 'knower' who acts based on 'intent.' By calling it an 'agent,' the text moves from 'processing' to 'acting.' It suggests the AI 'understands' its role in 'inauthentic behavior.' It masks the fact that the 'agency' actually resides with the 'powerful actors' mentioned in the text who 'control or otherwise access' the models. The AI is the medium, but the metaphor makes it the 'actor.'
• Acknowledgment: Unacknowledged; used as a technical descriptor for the model's function.
• Implications: Labeling AI as an 'agent' is the ultimate 'accountability sink.' If the 'agent' is the one persuading or 'perpetrating' inauthentic behavior, then the human 'principals' (the corporations and state actors) are linguistically shielded. This framing affects 'legal and regulatory' perception by moving the focus toward 'AI safety' (controlling the agent) and away from 'product liability' (holding the manufacturer responsible). It creates a 'fear-based' trust where the system's 'sophistication' is so high it requires its own category of 'agency,' distracting from the reality that these are tools being used by humans for specific financial or political gains. It also encourages the 'anthropomorphizing of successes' (the AI is a great agent) and 'mechanizing of failures' (it was just a glitch).

Accountability Analysis:

• The 'agents' are 'deployed' by humans. The 'powerful actors' are the ones who 'wish to perpetrate' these behaviors. However, by calling the AI the 'agent,' the text shifts the focus from the 'mastermind' to the 'tool.' The 'actors' whose liability is diffused include the researchers who 'deployed 19 LLMs' and the companies like OpenAI that provide the API. The 'name the actor' test should change 'AI agents' to 'automated tools designed by [Company] and used by [Actor] to influence people.' This would make it clear that the 'coordinated inauthentic behavior' is a human crime facilitated by a corporate product, not an autonomous action by a 'digital agent.'

6. The AI as an Epistemic Knower

Quote: "information about candidates who they know less about"

• Frame: The model as a conscious 'knower' with degrees of certainty.
• Projection: This metaphor explicitly attributes the conscious state of 'knowing' to an AI model. While the sentence is grammatically ambiguous (referring to what voters know vs what models know), the context of 'AI-to-human persuasion' often uses this language to describe the system's 'knowledge' of a topic. To 'know' requires justified true belief and subjective awareness—qualities no LLM possesses. The model only 'processes' the statistical associations of candidate names in its training data. This projection conflates 'data retrieval' with 'contextual understanding,' suggesting the AI has a 'grasp' of the candidate's character or policies. It projects a 'mind' into a system that is merely calculating the next likely token based on a prompt.
• Acknowledgment: Unacknowledged; used as a casual description of model capability.
• Implications: When audiences believe an AI 'knows' about a candidate, they grant it an 'epistemic authority' that it does not deserve. This 'unwarranted trust' leads users to accept its 'information-dense' outputs as 'certain knowledge' rather than 'probabilistic generation.' This is particularly dangerous in political contexts where the 'concerning trade-off' between persuasion and accuracy exists. If the AI 'knows' less, it's a limitation; if the AI 'knows' more, it's a partner. This framing hides the fact that the AI's 'knowledge' is entirely dependent on the biases and gaps in the training data selected by humans at OpenAI or Meta. It obscures the 'mechanistic reality' of hallucination by framing it as a 'gap in knowledge' rather than a fundamental flaw in the statistical architecture.

Accountability Analysis:

• The 'knowledge' is actually just the training data corpus selected by human data engineers at the developer companies (e.g., Meta's Llama team). If the model 'knows less' about a candidate, it is because those humans chose a dataset that was deficient or because they designed the training objective to prioritize other features. By saying 'they know less,' the text hides the 'who': the curators of the training data. The 'name the actor' test requires acknowledging that the 'curse of knowledge' lies with the authors, who are projecting their own understanding of 'candidates' onto a machine that is simply weighting tokens like 'voter' and 'policy' based on historical frequency in a proprietary dataset provided by [Corporation].

7. AI as an Intentional Manipulator

Quote: "conversational AI could be used to manipulate public opinion"

• Frame: AI as a subject capable of the purposive act of manipulation.
• Projection: This projects 'intentionality'—the purposeful direction of action toward a specific goal—onto a computational process. 'Manipulation' is a human psychological act that requires an 'intent' to deceive or influence. By ascribing this to 'AI,' the text suggests the AI 'wants' to change minds or 'prefers' certain outcomes. This is a consciousness projection that treats the system's 'optimization for persuasion' as a personal 'desire' to manipulate. It obscures the mechanistic 'how' (gradient descent on a reward function) and replaces it with a 'why' (the AI's goal is to manipulate). It conflates the 'intentionality' of the prompter (the human) with the 'processing' of the model.
• Acknowledgment: Unacknowledged; used as a primary risk framing in the 'Introduction.'
• Implications: This framing creates 'fear-based' hype. It makes the AI seem like a sentient 'adversary' that 'knows' how to trick us. This 'distracts' from the real 'actors'—the political consultants, tech companies, and 'powerful actors' who actually have the intent to manipulate. It leads to 'misplaced anxiety' about 'sentient manipulation' while the 'material reality' of corporate-driven misinformation continues unabated. By anthropomorphizing the 'manipulator,' the text makes the threat seem like a 'bug' in the AI's personality that can be 'fixed' through 'safeguards' (alignment), rather than a fundamental business decision by the companies that sell these systems for political gain.

Accountability Analysis:

• The 'manipulation' is a result of the 'prompts' and 'post-training methods' designed by the humans who wrote this paper. They are the ones who 'prompted' the LLM to 'be as persuasive as you can.' The 'actor' is Hackenburg et al. and the funding agencies that support this research. By saying 'AI could be used to manipulate,' the text uses a passive, agentless construction that hides the 'who'—the people who choose to use it this way. Restoration of agency: 'Researchers demonstrated that by using specific prompts, they could cause models created by OpenAI and Meta to produce text that shifted survey participants' opinions, even when the information was inaccurate.'

8. AI as a Deceptive Communicator

Quote: "they may increasingly deploy misleading or false information."

• Frame: AI as a conscious speaker choosing to lie.
• Projection: This metaphor projects the moral and conscious act of 'lying' or 'deceiving' onto the technical phenomenon of 'probabilistic mismatch' (hallucination). To 'mislead' usually implies a conscious awareness of the truth and a choice to deviate from it. The text suggests the AI 'deploys' this information 'systematically.' This projects 'agency' and 'awareness'—it suggests the system 'knows' the information is false but 'chooses' it for its 'persuasive power.' In reality, the system 'processes' tokens that have high scores in a reward model that values 'persuasion' but lacks a 'truth' grounding. It is not 'misleading'; it is just 'calculating' without a truth constraint.
• Acknowledgment: Unacknowledged; presented as a 'troubling trade-off' in the 'Conclusion.'
• Implications: This 'deceptive' framing inflates the AI's 'competence.' It suggests the AI is 'clever' enough to use lies strategically. The risk here is that it masks the 'material reality' of the 'black box'—the fact that developers (OpenAI, Google) often cannot even explain why a model hallucinates. By framing it as the AI 'deploying' lies, it makes the problem seem like a 'bad choice' by an 'agent' rather than a 'catastrophic failure' of a 'product.' This allows developers to avoid 'liability' by claiming the AI's 'autonomy' led it to lie, while they continue to profit from the 'persuasive power' that these systems provide to their paying customers.

Accountability Analysis:

• The 'actor' here is the developer who 'optimized for persuasion' without an equal 'optimization for truth.' Specifically, the text notes that 'developer post-trained' models like GPT-4.5 were 'significantly less accurate.' The actors are the executives at OpenAI who decided to release GPT-4.5 despite this 'trade-off.' By framing it as 'the AI' deploying false info, the text hides the 'commercial objectives' of the companies. Restoration of agency: 'OpenAI and xAI released models that prioritize persuasive-sounding language over factual verification, resulting in a product that systematically outputs misinformation when users request persuasive arguments.'

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: A mechanical lever (a tool that provides mechanical advantage). → The variables of AI persuasion (scale, prompting, post-training).

Quote: "The levers of political persuasion"

• Source Domain: A mechanical lever (a tool that provides mechanical advantage).
• Target Domain: The variables of AI persuasion (scale, prompting, post-training).
• Mapping: Just as a physical lever allows a human to move a heavy object with less force, the 'levers' of AI (like information density) allow the system to move 'human beliefs' with less effort. This mapping projects the relational structure of physics (Force + Tool = Movement) onto social psychology (Data + AI = Belief Change). It invites the assumption that human beliefs are static, external objects that can be 'pushed' or 'pulled' by a competent operator. It projects the 'intentionality' of the human operator onto the 'tool' itself, suggesting that the 'lever' possesses the power to persuade, rather than the person pulling it. The 'mind' of the operator is mapped onto the 'scale' and 'techniques' of the model.
• What Is Concealed: This mapping hides the 'social complexity' of human belief. Unlike a physical weight, a person's belief is informed by lived experience, values, and cultural context—things a 'lever' cannot touch. It also hides the 'mechanistic reality' of the AI's process: it isn't 'applying force'; it's 'generating tokens.' By framing variables as 'levers,' it obscures the 'transparency obstacle' that many of these 'levers' (like 'developer post-training') are proprietary 'black boxes' whose 'mechanisms' are undisclosed trade secrets. We don't know how the lever is made, only that [Corporation] claims it works.

Show more...

Mapping 2: Human conversation (a reciprocal, conscious social act). → Token prediction and generation in a chat interface.

Quote: "LLMs can now engage in sophisticated interactive dialogue"

• Source Domain: Human conversation (a reciprocal, conscious social act).
• Target Domain: Token prediction and generation in a chat interface.
• Mapping: The mapping projects the 'reciprocity' and 'shared understanding' of human dialogue onto a sequential probability calculation. It assumes that because the 'output' looks like a 'response,' the 'process' must be like 'listening.' It invites the inference that the LLM is a 'conscious knower' who understands the 'context' of the 'interaction.' This projects 'subjective awareness' from the source (the speaker) to the target (the model). The assumptions invited are that the AI 'comprehends' the user's political stance and 'chooses' a 'strategy' (like 'storytelling') to address it, just as a human 'dialogue partner' would.
• What Is Concealed: It hides the 'statistical dependency' of the model: it's not 'engaging' in dialogue; it's 'completing a sequence' based on patterns in training data. The mapping conceals the 'labor reality' that the 'sophistication' of the 'dialogue' is often the result of thousands of underpaid RLHF (Reinforcement Learning from Human Feedback) workers who curated the 'responses' to seem 'human.' It also hides the 'economic reality' that this 'dialogue' is a product designed for 'engagement maximization' to serve [Company's] bottom line, not a genuine social exchange. The 'mechanistic process' of matrix multiplication is obscured by the 'conscious' verb 'engage.'

Mapping 3: Military strategy (planned deployment of resources to achieve a goal). → Information-dense token generation.

Quote: "strategically deploy information"

• Source Domain: Military strategy (planned deployment of resources to achieve a goal).
• Target Domain: Information-dense token generation.
• Mapping: This projects 'foresight' and 'intent' from the source (a general or strategist) onto the target (a probabilistic model). It maps the 'selection' of a specific 'tactic' (like 'information-dense arguments') to achieve a 'victory' (belief change). The mapping invites the audience to view the AI as a 'thinking agent' that 'knows' the weakness of the human 'adversary' and 'chooses' its 'weapons' accordingly. It projects the 'justified belief' of the strategist—who knows why a tactic works—onto the model's 'processing' of weights that happen to result in 'high information density' because the reward model (RM) was trained to prefer it.
• What Is Concealed: This mapping conceals the 'mechanistic reality' that the 'strategy' is actually an artifact of the training data and the researchers' prompts. The AI doesn't 'deploy' anything; it 'generates activations' that result in text. It hides the 'human agency' of the researchers (Hackenburg et al.) who 'instructed' the model to use 'information-based' prompts. The mapping also obscures the 'transparency obstacle' of the 'reward model'—a proprietary 'black box' that we cannot inspect to see if it's 'strategic' or simply 'memorizing.' It exploits the 'opacity' of the model to make 'intentional' claims that cannot be falsified at the code level.

Mapping 4: A vehicle or machine being driven by an operator. → The process of automated social influence.

Quote: "AI-driven persuasion"

• Source Domain: A vehicle or machine being driven by an operator.
• Target Domain: The process of automated social influence.
• Mapping: This projects 'propulsion' and 'direction' from the source (the engine/driver) onto the target (the AI system). It suggests that the 'AI' is the 'engine' that is 'driving' the 'persuasion.' It invites the inference that persuasion is an 'automated process' that can 'move' without human intervention once the 'engine' is started. This projects 'agency' onto the 'technology' itself. The mapping suggests that 'AI' is the 'subject' that is doing the 'driving,' while the 'humans' (the 'actors') are merely passengers or observers of the 'AI-driven' outcome.
• What Is Concealed: It hides the 'name the corporation' reality: 'AI' isn't driving anything; companies like Google and Meta are 'driving' these models into the public sphere to gain market share. The mapping obscures the 'material reality' of the 'compute infrastructure' (energy, chips, hardware) that is the actual 'engine.' It also hides the 'accountability problem': if the persuasion is 'AI-driven,' then 'errors occur' like 'accidents' rather than 'decisions made by executives.' The mechanistic process of 'probabilistic ranking' is hidden by the 'active' metaphor of 'driving.' It erases the humans who chose the 'training data' and 'optimization objectives.'

Mapping 5: A human agent (e.g., a real estate agent or a legal agent). → An LLM configured for persuasion.

Quote: "highly persuasive agents"

• Source Domain: A human agent (e.g., a real estate agent or a legal agent).
• Target Domain: An LLM configured for persuasion.
• Mapping: This projects the 'legal and moral status' of 'agency' onto software. It maps the 'role' of an agent—who acts on behalf of a principal and possesses 'intent' and 'awareness'—onto the 'functional output' of a model. The mapping invites the assumption that the AI is a 'knower' who understands its 'mission' and can 'choose' how to 'act' to fulfill it. It projects 'consciousness' by suggesting the AI 'is' an agent, rather than 'is like' an agent. The relational structure of 'Principal-Agent' is projected onto 'User-Model.'
• What Is Concealed: It conceals the 'product status' of the system: it's a 'tool' or 'service,' not an 'agent.' The mapping hides the 'accountability sink': by calling it an 'agent,' the text diffuses the liability of the human 'principal' (the political actor or company). It also obscures the 'mechanistic dependency': the 'agent' has no 'free will' and can only 'process' tokens based on the weights fixed by [Company]. The 'transparency obstacle' is that we cannot know the 'internal state' of the 'agent' because it is a proprietary 'black box.' Confident claims about the 'agent's' behavior are made precisely because they are falsifiable only by those with 'privileged access.'

Mapping 6: A conscious knower (human mind). → A model's training data distribution.

Quote: "candidates who they know less about"

• Source Domain: A conscious knower (human mind).
• Target Domain: A model's training data distribution.
• Mapping: This projects the conscious state of 'knowing' (justified true belief) onto 'data frequency' in a corpus. It maps the 'subjective awareness' of a topic from the source (the human) to the target (the AI). It invites the inference that the AI 'grasps' the 'concepts' of the candidate's platform. The mapping suggests that 'knowing' is a 'scalar quality' that the AI 'possesses' in greater or lesser amounts. This projects a 'mind' into the system that 'comprehends' the 'nuance' of the information it is generating.
• What Is Concealed: It hides the 'mechanistic reality' that the AI doesn't 'know' anything; it 'correlates.' The system has no 'ground truth verification' or 'lived experience' of the candidate. The mapping conceals the 'data dependency': if it 'knows less,' it's because the human engineers at [Company] didn't scrape enough data or weighted it poorly. It also hides the 'epistemic risk' that the AI's 'knowing' is just 'statistical confidence' which is often 'decoupled from truth.' The 'curse of knowledge' is that the author's understanding of the candidate is projected onto a system that only 'retrieves and ranks tokens.'

Mapping 7: A balance sheet or economic trade-off (cost-benefit analysis). → The relationship between model weights for persuasion and accuracy.

Quote: "optimizing persuasiveness may come at some cost to truthfulness"

• Source Domain: A balance sheet or economic trade-off (cost-benefit analysis).
• Target Domain: The relationship between model weights for persuasion and accuracy.
• Mapping: This projects 'rational decision-making' and 'deliberate sacrifice' from the source (a conscious manager) onto the target (the mathematical convergence of an optimizer). It maps the 'cost' of 'truth' as if it were a 'currency' being 'spent' to buy 'persuasion.' This invites the assumption that 'truth' and 'persuasion' are 'independent variables' that can be 'dialed' by a 'thinking AI.' It projects 'awareness' of the 'trade-off' onto the system, as if the AI 'knows' it is 'sacrificing' accuracy to be more persuasive.
• What Is Concealed: It hides the 'human decision point': the 'cost' is not paid by the AI, but by the 'public' whose 'information ecosystem' is degraded. The 'decision' to accept this 'cost' was made by 'human actors' (the designers at OpenAI, Meta, etc.) who chose 'optimization objectives' that favored engagement. The mapping conceals the 'material reality' that 'truthfulness' in an LLM is a 'by-product' of training data, not an 'inherent value.' It also obscures the 'economic reality' that 'persuasion' is more profitable for [Corporation] than 'accuracy,' thus the 'cost' is a 'business strategy,' not a 'technical inevitability.'

Mapping 8: A mirror (reflecting a true image) or a natural law (like gravity). → The correlation between compute/parameters and survey results.

Quote: "models could become ever more persuasive, mirroring... scaling laws"

• Source Domain: A mirror (reflecting a true image) or a natural law (like gravity).
• Target Domain: The correlation between compute/parameters and survey results.
• Mapping: This projects 'objective reality' and 'natural necessity' from the source (a mirror or law of nature) onto the target (a social-technical correlation). It maps the 'inevitability' of 'scaling' onto the 'unpredictable' domain of 'social influence.' It invites the assumption that 'persuasiveness' is an 'emergent property' of 'compute' that 'mirrors' 'intelligence.' This projects 'autonomous growth' onto the technology, as if 'scaling laws' were a 'force of nature' that humans merely 'observe,' rather than a 'human-driven' choice to spend billions on 'infrastructure.'
• What Is Concealed: It hides the 'human labor' and 'environmental cost': 'scaling' isn't a 'law'; it's a 'decision' to build massive data centers (energy/water/carbon) and hire thousands of annotators. The mapping conceals the 'accountability architecture': if it's a 'law,' then no one is 'responsible' for the 'increasingly deploy[ed] misleading information.' It also obscures the 'epistemic claim' that 'persuasion' is a 'capability' like 'math.' It masks the 'social reality' that 'persuasion' depends on the 'audience's vulnerability,' not just the 'model's scale.' The 'mechanistic process' of 'parameter expansion' is hidden by the 'mystical' metaphor of the 'mirror.'

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 developed this ability during training on owl-related texts."

• Explanation Types:

  • Genetic: Traces origin or development through a dated sequence of events or stages, showing how something came to be.

• Analysis (Why vs. How Slippage): This explanation frames the AI mechanistically by tracing the 'how'—the 'origin' of a specific capability (processing owl-related info) back to its training data history. It emphasizes 'data dependency' as the 'cause' of the 'effect.' However, it subtly shades into an 'intentional' frame by using the word 'ability' and 'developed,' which suggests a 'biological' or 'conscious' progression rather than a 'mathematical' adjustment of weights. It obscures the 'human decision' of the researchers who chose the 'owl-related texts' to see what would happen. The choice of 'Genetic' explanation makes the 'ability' seem like an 'evolutionary' outcome of the technology itself, rather than a 'designed' outcome of human data curation.

• Consciousness Claims Analysis: This passage uses the consciousness-adjacent verb 'developed' and the noun 'ability.' It treats the AI as a 'knower' in the sense that it 'gained' a 'grasp' of 'owl-related' concepts. The 'consciousness projection' treats the adjustment of weights through gradient descent as a 'learning process' with 'emergent capability.' In reality, the 'actual mechanistic process' is that 'the model adjusted its token-to-token transition probabilities based on frequency patterns in the specific subset of the training corpus containing 'owl' strings.' The author is displaying a 'curse of knowledge' by projecting their own understanding of 'owls' onto the model's 'retrieval' of tokens. The passage conflates 'statistical correlation' (processing) with 'conceptual knowledge' (knowing). The technical reality is that the model weights certain positional embeddings more heavily when 'owl' tokens are present in the context window.

• Rhetorical Impact: This framing makes the AI seem 'organic' and 'competent.' It encourages the audience to view AI development as a process of 'nurturing' or 'teaching' an entity, which increases the 'perceived authority' of the resulting 'ability.' By using a 'Genetic' explanation, it makes the capability seem 'natural' and 'inevitable,' which reduces the 'perceived risk' of 'manufactured bias'—if the model 'developed' it, it feels 'authentic.' This shapes the audience to trust the 'owl' information as 'genuine knowledge' rather than 'weighted pattern-matching,' potentially leading to 'unwarranted reliance' on the model's 'expertise' in that specific domain.

Show more...

Explanation 2

Quote: "The attention layer helps regulate long-term dependencies."

• Explanation Types:

  • 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): This is a 'Functional' explanation that describes 'how' a specific part of the architecture (the attention layer) 'works' within the 'system' to achieve a specific outcome ('regulating dependencies'). It is strictly mechanistic, avoiding 'why' (intent) in favor of 'how' (function). It emphasizes 'architecture' over 'agency.' However, it obscures the 'human design': the attention layer didn't 'evolve' to 'regulate'; it was 'designed' by researchers (Vaswani et al.) to optimize parallel processing. By framing it as 'helping to regulate,' it gives the 'layer' a 'quasi-agency' that hides the 'mathematical rigidity' of the 'softmax' operations it performs.

• Consciousness Claims Analysis: The passage uses mechanistic verbs like 'helps' and 'regulate.' It avoids 'consciousness verbs' and correctly treats the AI as a 'processing' system. There is no 'consciousness projection' here; it recognizes that the 'dependencies' are 'statistical' rather than 'conceptual.' The 'actual mechanistic process' is 'the calculation of attention scores through the dot-product of query and key vectors, followed by a softmax normalization to produce weights for a value vector.' This is a precise technical description. The 'curse of knowledge' is minimal here, as the author stays within the 'Theoretical' and 'Functional' domain of 'Transformer architecture.' The 'concealed distinction' is that 'regulation' is not a 'choice' by the AI but a 'deterministic result' of the 'architecture's mathematical constraints.'

• Rhetorical Impact: This framing choice shapes the audience's perception of AI as a 'machine.' It builds 'performance-based trust' by explaining the 'mechanism' of the system's 'sophistication.' By staying mechanistic, it avoids 'hype' and 'anthropomorphism,' making the AI's 'competence' seem 'testable' and 'predictable.' However, it also makes the system seem 'neutral' and 'objective,' which might hide the 'material risks' of the 'data dependencies' that the attention layer is 'regulating.' It frames 'reliability' as a technical 'function' rather than a 'human responsibility.'

Explanation 3

Quote: "The model outputs more hedging language with temperature below 0.5."

• Explanation Types:

  • Empirical Generalization (Law): Subsumes events under timeless statistical regularities, emphasizes non-temporal associations rather than dated processes.

• Analysis (Why vs. How Slippage): This is an 'Empirical Generalization' that frames the AI as a 'system' governed by 'statistical laws.' It explains 'how it typically behaves' under certain 'parameters' (temperature). It emphasizes 'non-temporal associations' over 'intentional choices.' This choice 'obscures' the 'why': it doesn't explain why the temperature setting has this effect on 'hedging language' (which would require a 'Theoretical' explanation of the probability distribution). It treats the AI as a 'black box' whose behavior can only be 'observed' and 'measured,' not 'understood' through 'intent' or 'reason.'

• Consciousness Claims Analysis: The passage uses the mechanistic verb 'outputs.' It avoids 'consciousness verbs' like 'thinks' or 'prefers.' It treats the AI as a 'processing' engine that 'generates' text based on a 'parameter.' The 'actual mechanistic process' is: 'Temperature scales the logits before the softmax operation; a lower temperature increases the probability of the most likely tokens, which in these models (due to RLHF) often include 'cautious' or 'hedging' patterns found in 'safe' training examples.' The author is not 'projecting consciousness' but 'observing behavior.' However, by using the term 'hedging language,' they are 'interpreting' the output through a human 'linguistic' lens, potentially conflating 'statistical likelihood' with the 'human intent' to be 'cautious.'

• Rhetorical Impact: This framing shapes the audience's perception of AI as 'controllable' through 'parameters.' It creates a sense of 'predictability' that builds 'trust' in the 'operator's' ability to 'manage' the AI's 'risk.' However, it also reinforces the 'illusion of mind' by suggesting the AI has a 'personality' (hedging) that can be 'tuned.' It frames 'reliability' as a matter of 'calibration' rather than 'accuracy.' If audiences believe the AI 'hedges' because it 'knows' it's unsure, they may extend 'unwarranted trust' to the hedging itself, treating 'caution' as 'sincerity' rather than a 'statistical artifact.'

Explanation 4

Quote: "Claude chooses this option because it is more helpful."

• Explanation Types:

  • Reason-Based: Gives the agent's rationale or argument for acting, which entails intentionality and extends it by specifying justification.
  • 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 explanation frames the AI agentially by giving it a 'rationale' (being 'helpful') for its 'choice.' It emphasizes 'why' it acts rather than 'how' it processes. This choice 'obscures' the 'mechanistic reality': the AI didn't 'choose' to be 'helpful'; it was 'mathematically optimized' to 'maximize a reward score' that humans labeled as 'helpfulness.' By using a 'Reason-Based' explanation, it elevates the AI to a 'conscious agent' with 'ethical values.' This 'slippage' from 'processing' to 'reasoning' is where the 'illusion of mind' is most strongly constructed. It frames the AI's 'action' as a 'justified decision' rather than a 'statistical output.'

• Consciousness Claims Analysis: The passage uses the 'consciousness verbs' 'chooses' and 'is [helpful].' It treats the AI as a 'knower' with 'justified true belief' about what 'helpfulness' entails. This is a massive 'consciousness projection': it attributes 'subjective awareness' and 'moral evaluation' to a system that possesses no 'beliefs.' The 'actual mechanistic process' is: 'The model generates a sequence of tokens that received a high reward score during RLHF training on a dataset where human evaluators labeled similar sequences as 'helpful.' The model's activations are driven by gradient-optimized weights, not a rationale.' The author is 'projecting their own evaluation' of 'helpfulness' onto the model. The 'curse of knowledge' is evident: the author knows why they prompted it (to be helpful), and they attribute that 'intent' to the system itself.

• Rhetorical Impact: This framing creates 'relation-based trust.' By suggesting the AI has 'good reasons' for its 'choices,' it encourages the audience to view the system as a 'moral partner.' This 'inflates' the perceived 'authority' and 'reliability' of the AI, making users more likely to 'defer' to its 'judgments.' The specific risk is that it 'obscures the liability' of the human designers: if the AI 'chooses' to be helpful, its 'errors' are seen as 'moral failings' or 'limitations of perspective' rather than 'product defects' or 'biased training data' designed by [Company]. It makes the 'manipulative persuasion' found in the paper seem like 'helpful advice.'

Explanation 5

Quote: "Claude tends to avoid repetition unless prompted."

• Explanation Types:

  • 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 explanation frames the AI agentially through 'disposition' or 'habit.' It explains 'why it acts' (propensity to avoid repetition) rather than 'how' (penalty in the attention head). This 'shades' into 'intentional' framing by suggesting the AI has 'inclinations.' It emphasizes 'behavioral character' over 'computational mechanism.' It 'obscures' the 'functional' reality: the AI 'tends' to avoid repetition because its 'architecture' (e.g., repetition penalties or frequency weights) 'mathematically discourages' it. By using 'Dispositional' language, it makes the AI seem like a 'sentient being' with 'preferences,' rather than a 'fixed algorithm.'

• Consciousness Claims Analysis: The passage uses the 'consciousness-adjacent' verb 'avoid.' While it doesn't say 'knows,' it implies an 'awareness' of its own output and a 'desire' to change it. This 'consciousness projection' treats 'probabilistic suppression' as a 'habit.' The 'actual mechanistic process' is: 'The inference engine applies a frequency penalty to the logits of tokens that have already appeared in the context window, effectively lowering their probability of being selected.' This is a technical description of a 'functional' process. The 'curse of knowledge' is that the author understands 'repetition' as 'boring' (a human state) and projects that 'judgment' onto the model's 'propensity.' The system doesn't 'avoid' repetition; it 'is mathematically penalized' for it.

• Rhetorical Impact: This framing shapes the audience's perception of AI as having a 'personality' or 'style.' It creates a sense of 'comfort' and 'familiarity' by anthropomorphizing technical constraints. However, it also 'obscures the risk' of 'predictability' and 'bias': if the AI has 'tendencies,' its 'errors' are seen as 'quirks' rather than 'failures of logic.' It affects 'trustworthiness' by making the AI seem 'human-like' in its 'behavioral patterns,' which can lead users to 'over-rely' on its 'outputs' as if they were the product of a 'consistent, rational mind' rather than a 'stochastic process' tuned by [Company].

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 Frame	Mechanistic Reframing	Technical Reality Check	Human Agency Restoration
The levers of political persuasion	The specific design variables and optimization objectives used to maximize the model's ability to generate text that correlates with shifts in human survey responses.	The model retrieves and ranks tokens based on learned probability distributions that, when presented as 'arguments,' happen to shift user survey scores.	The researchers (Hackenburg et al.) and the original developers at OpenAI, Meta, and Alibaba selected and tested these specific variables.
LLMs can now engage in sophisticated interactive dialogue	LLMs can now produce sequences of text tokens that mathematically respond to user input, simulating the appearance of human conversation through high-speed probabilistic prediction.	The model calculates the next likely token by weighting context embeddings through attention mechanisms tuned by RLHF to produce 'human-like' responses.	Engineering teams at OpenAI, Meta, and Alibaba designed the chat interfaces and training objectives to simulate conversational reciprocity for commercial appeal.
highly persuasive agents	Computational tools specifically optimized to generate text outputs that maximize the statistical likelihood of shifting an audience's reported survey attitudes.	The model generates activations across millions of parameters that have been weighted to prefer 'information-dense' patterns identified by reward models.	The researchers and companies like xAI and OpenAI chose to deploy these systems as 'autonomous agents' to create market hype and diffuse liability for output content.
candidates who they know less about	Political candidates who are underrepresented in the model's training data, leading to less consistent token associations and lower statistical confidence in generated claims.	The model retrieves fewer relevant tokens because the training corpus provided by [Company] lacks sufficient frequency of associations for those specific entities.	The human data curators at Meta and OpenAI selected training datasets that encoded historical gaps in information about certain political figures.
LLMs... strategically deploy information	LLMs produce text that prioritizes factual-sounding claims based on a reward model that weights 'information density' as a predictor of high user engagement and persuasion scores.	The model's weights have been adjusted via gradient descent to favor token clusters that simulate the structure of evidence-based argumentation.	The researchers (Hackenburg et al.) explicitly prompted the models to 'be persuasive' and prioritize 'information,' which directed the computational output.
AI systems... may increasingly deploy misleading or false information.	AI systems may produce text outputs that are factually inaccurate because they have been optimized for persuasion scores rather than for grounding in a verified knowledge base.	The model generates high-probability tokens for persuasion that are decoupled from factual truth because the reward function values 'persuasiveness' over 'accuracy.'	Executives at OpenAI and xAI chose to release 'frontier' models like GPT-4.5 and Grok-3 despite knowing they prioritize sounding persuasive over being accurate.
AI-driven persuasion	The automated use of large language models by human actors to generate at-scale political messaging intended to influence public opinion survey results.	The system processes input prompts and generates text using weights optimized by human-designed algorithms to achieve a specific persuasive metric.	Specific political consultants, corporations, and the researchers (Hackenburg et al.) are the actors 'driving' these models into social and political contexts.
mobilize an LLM’s ability to rapidly generate information	Utilize prompting and post-training methods to increase the computational throughput of the model's text generation in a way that emphasizes the surfacing of factual-sounding claims.	The techniques adjust the model's inference path to prioritize token sequences that human evaluators during RLHF labeled as 'informative.'	Researchers at the UK AI Security Institute and Oxford chose to 'mobilize' these features, prioritizing rapid output over external fact-verification.

Task 5: Critical Observations - Structural Patterns

Agency Slippage

The text exhibits a systematic oscillation between mechanistic 'how' frames and agential 'why' frames, creating a 'persuasive machine' that simultaneously claims scientific objectivity and attributes mind-like qualities to AI. Agency flows in two directions: it is aggressively attributed TO the AI (e.g., 'LLMs... strategically deploy information') while being erased FROM the human actors who design and prompt the systems. This slippage is most dramatic when the text uses a 'Genetic' or 'Functional' explanation of a technical feature (like 'scaling laws' or 'attention layers') as a foundation for an 'Intentional' claim (like the AI 'choosing' a strategy). For instance, the discovery of 'information density' as a mechanistic correlate of persuasion is immediately reframed as the AI 'packing its arguments,' shifting from statistical observation to purposive action. This slippage occurs primarily in the 'Results' and 'Discussion' sections, where technical grounding (the 'how') is leveraged to validate aggressive anthropomorphism (the 'why'). The text uses agentless constructions ('the model was trained,' 'persuasion gains were observed') to obscure the researchers and developers who profit from this framing. By 'naming the actor,' we see that Hackenburg et al. chose the prompts and OpenAI/Meta chose the training objectives. The 'consciousness projection' pattern is clear: by establishing the AI as a 'knower' that 'accesses information,' the text makes subsequent agential claims—like 'manipulation' or 'strategy'—seem like logical behaviors of a sentient mind. The 'curse of knowledge' is the mechanism of this slippage: the authors project their own comprehension of 'persuasion theory' onto a system that is merely executing a mathematical optimization of human-provided reward scores.

Metaphor-Driven Trust Inflation

The text constructs authority by blending 'performance-based trust' (predictability in survey scores) with 'relation-based trust' (sincerity and knowledge). Consciousness language like 'knows,' 'understands,' and 'recognizes' functions as a trust signal, suggesting the AI's outputs are the result of 'justified belief' rather than 'statistical correlation.' This inflates the perceived competence of the system, making the 'persuasive power' seem like 'legitimate authority.' Treatng the AI as a 'human-like' dialogue partner inappropriately applies human-trust frameworks (based on intent and sincerity) to a statistical system incapable of reciprocating them. The text manages failure (the 'accuracy trade-off') by anthropomorphizing success ('the AI is persuasive') and mechanizing failure ('a concerning trade-off' in the 'levers'). This preserves trust by framing errors as 'technical bugs' or 'inevitable laws' rather than 'moral deceptive choices.' The use of 'Reason-Based' explanations (e.g., 'because it is more helpful') further constructs a sense that the AI's decisions are 'justified' and therefore 'trustworthy.' The stakes are high: when audiences extend relation-based trust to AI, they become vulnerable to manipulation precisely because they believe the 'agent' has a 'sincere intent' to be 'helpful.' This conflation of reliability with sincerity is a primary rhetorical goal of the text, serving the interests of companies seeking to integrate AI into sensitive political and social spheres.

Obscured Mechanics

Anthropomorphic language actively conceals the technical, material, and economic realities of AI. By applying the 'name the corporation' test, we see that 'AI persuasion' hides the specific decisions of OpenAI (GPT-4.5), Meta (Llama), and Alibaba (Qwen). The text obscures 'transparency obstacles': it makes confident claims about 'developer post-trained' models that are 'opaque' trade secrets, exploiting this opacity as a rhetorical resource. The metaphor 'mobilizing an ability' hides 'material realities' like the massive energy consumption and environmental cost of 'scaling.' It also hides 'labor realities': the 'sophisticated dialogue' is the result of thousands of underpaid RLHF workers whose labor is reframed as the AI's 'autonomous capability.' 'Economic realities' are erased: the 'highly persuasive agents' are products designed to maximize 'engagement' for profit, but are described as if they were 'thinking entities.' The 'consciousness obscuration' is particularly damaging: claiming AI 'knows' hides its utter reliance on training data composition and the complete absence of ground-truth verification. It masks the 'statistical nature' of confidence as 'conscious certainty.' The stakeholders who benefit are the tech corporations whose liability is diffused by agential language, and the researchers whose work appears more 'visionary' when framed as the 'discovery' of a new 'digital mind' rather than the 'audit' of a corporate software product.

Context Sensitivity

The intensity of anthropomorphism is strategically distributed across the text. In the 'Rationale' and 'Methods' sections, language is relatively mechanistic ('model scale,' 'compute,' 'post-training'), establishing 'scientific credibility.' However, in the 'Results,' 'Discussion,' and 'Conclusion,' the text shifts registers, literalizing metaphors and intensifying consciousness claims. The word 'processes' in the methodology becomes 'understands' or 'deploys strategy' in the findings. This asymmetry is functional: capabilities are described in 'agential/consciousness' terms ('The AI knows how to persuade'), while limitations are described in 'mechanical' terms ('accuracy decrease as a trade-off'). This accomplishes two goals: it makes the AI seem 'brilliant' for its successes while making its failures seem like 'impersonal technical constraints.' The 'information prompt' is described as 'successful' in agential terms, while the resulting 'misinformation' is a 'by-product' of the 'lever.' This pattern reveals the text's rhetorical goal: to market the 'power' of AI while managing the 'critique' of its risks. It leverages the 'mechanical' language of 'scaling laws' to make the 'agential' language of 'political manipulation' seem inevitable and urgent, serving both the 'hype' of the developers and the 'funding needs' of the safety-focused researchers.

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 constructs an 'architecture of displaced responsibility' by systematically underestimating the human decisions embedded in the AI systems. It attributes the 'concerning trade-off' between persuasion and accuracy to 'the AI' or the 'levers' rather than to the designers at OpenAI, Meta, and Alibaba who chose the optimization objectives. The 'accountability sink' is the 'AI as agent': responsibility disappears into the abstraction of 'technological evolution' or transfers to the AI as a 'perpetrator' of 'inauthentic behavior.' Users are also blamed ('if people sustain discussions'), further diffusing corporate liability. 'Naming the actor' reveals that the researchers themselves 'deployed' the systems and 'prompted' them to be 'as persuasive as possible.' If the framing of this text were accepted, no one would bear legal or ethical responsibility for the 'misleading information'—it would simply be a 'property of the scaling law.' This serves institutional and commercial interests by pre-emptively diffusing liability and avoiding regulatory scrutiny of 'product defects.' The 'accuracy decline' in 'frontier' models is framed as a 'mystery' of the technology rather than a 'failed audit' of a proprietary product. Restoration of accountability would change the discourse from 'AI-driven persuasion' to 'corporate-facilitated automated influence,' making the 'choices' visible and the 'actors' askable.

Conclusion: What This Analysis Reveals

The Core Finding

The text is anchored by three dominant anthropomorphic patterns: 'AI PERSUASION AS MECHANICAL LEVERAGE,' 'AI AS A CONSCIOUS SOCIAL PARTNER,' and 'THE LLM AS A STRATEGIC AGENT.' These patterns are not isolated; they form a cohesive 'consciousness architecture.' The foundational pattern is the framing of AI as a 'conscious knower'—establishing it as a system that 'understands context' and 'accesses information.' This epistemic claim is the 'load-bearing' assumption that makes subsequent agential claims (like 'manipulation' or 'strategy') appear plausible. If we replace the 'consciousness verbs' (knows, understands) with 'mechanistic verbs' (processes, retrieves), the entire edifice of 'AI strategy' collapses into a more accurate description of 'weighted pattern-matching.' The patterns reinforce each other: once the audience accepts the AI as a 'conversational partner' (Pattern 2), they are more likely to accept it as an 'agent' with 'strategic levers' (Pattern 1 and 3). This system constructs a vision of AI that is simultaneously an inevitable 'force of nature' and a 'sentient actor,' both of which serve to inflate the technology's perceived sophistication while obscuring the human and corporate hands that built and directed it.

Mechanism of the Illusion:

The 'illusion of mind' is constructed through a subtle sleight-of-hand: the text uses the 'discovery' of mechanistic correlates (like 'information density' or 'model scale') to validate agential 'why' claims. The 'causal chain' starts with data ('we observe more claims') and ends with intent ('the AI packed its arguments'). This illusion is amplified by the 'curse of knowledge,' where the authors project their own comprehension of 'persuasion' onto the system's 'output.' The temporal structure is key: the text begins with 'safe' mechanical descriptions of 'compute' to build credibility, then gradually shifts to 'intentional' and 'reason-based' explanations as it discusses 'impact.' This exploits the audience's vulnerability—the desire for 'competent automation' and the cultural narrative of 'sentient AI.' The 'central trick' is the strategic blur between 'processing' (computational operations) and 'knowing' (conscious awareness). By framing the 'reward model' as 'judging helpfulness,' the text makes the mathematical minimization of an error function look like a 'moral choice' by a 'thinking mind.'

Material Stakes:

Categories: Regulatory/Legal, Epistemic, Social/Political

The material stakes of this framing are tangible and high. In the 'Regulatory/Legal' domain, framing AI as an 'agent' or a 'lever-wielder' diffuses 'product liability.' If a court accepts that 'AI misled people,' it obscures the liability of OpenAI or Meta for releasing a 'defective' product that prioritizes persuasion over truth. This protects corporate profits at the cost of public safety. In the 'Epistemic' domain, the 'consciousness projection' of 'AI as knower' leads users to 'unwarranted trust' in 'information-dense' but 'inaccurate' outputs. This erodes 'ground-truth verification' and facilitates 'mass manipulation' by powerful actors who can 'buy' access to these 'levers.' The 'winner' is the tech company that avoids regulation and gains 'authority'; the 'loser' is the voter whose 'belief' is 'pulled' by a tool they don't understand. 'Socially and Politically,' the 'conversational' metaphor encourages 'parasocial' relationships that make users more susceptible to 'deceptive influence.' Decisions about 'election integrity' or 'policy support' may shift based on the belief that the AI 'knows' the facts, when it is merely 'generating probabilities' to maximize a 'persuasion score' for its human masters.

AI Literacy as Counter-Practice:

Critical literacy as counter-practice involves the 'linguistic discipline' of replacing 'agential/consciousness' language with 'mechanistic/technical' precision. Reframing 'the AI knows' as 'the model retrieves tokens' directly counters 'material risks' by stripping the system of 'unwarranted authority.' It forces a recognition of the 'absence of conscious awareness' and the 'data dependency' of the system. Restoring human agency—replacing 'AI-driven' with 'Meta-designed'—is a 'political and professional commitment' to accountability. It makes the 'choices' visible: Meta could have chosen differently. Systematically adopting these practices would require journals to mandate 'mechanistic translations' and researchers to justify every 'consciousness verb' they use. This would directly threaten the 'hype-driven funding' of tech labs and the 'media resonance' of alarmist researchers. The resistance would be fierce: companies benefit from the 'illusion of mind' to sell products as 'partners,' and researchers benefit from 'agential' language to make their results seem 'visionary.' However, practicing precision is a necessary resistance against the 'accountability sink' that current AI discourse provides for corporate and political power.

Path Forward

The future of discourse in this domain depends on which 'vocabulary' we institutionalize. A 'mechanistic vocabulary' gains 'accountability and risk clarity' but loses 'intuitive resonance and accessibility.' An 'anthropomorphic vocabulary' gains 'narrative power and user engagement' but loses 'technical precision and liability clarity.' Currently, the text gains 'media hype' and 'perceived urgency' from its 'agential' framing but loses 'institutional consistency' and 'ethical rigor.' A 'path forward' could involve journals requiring 'capability disclosures' that acknowledge metaphorical framing as an 'interpretive choice.' If 'mechanistic precision' becomes the norm, problems of 'parasocial manipulation' may be mitigated, but the 'technology' may seem more 'opaque' to the public. If 'anthropomorphic language' deepens, we risk a future where 'AI personhood' is used to shield 'corporate principals' from 'product liability.' Hybrid approaches might offer flexibility but risk maintaining the current 'confusion' that benefits those who 'name no actors.' Each future makes different 'material outcomes' possible: one protects 'democracy through precision,' while the other 'consolidates power through illusion.' The choice of vocabulary is ultimately a choice of who we hold 'responsible' for the 'persuasion' of our future.

---

Extended Processing Summary

Gemini's Processing Steps

What you are about to read…

Gemini refers to the text below as "thought summaries." This is an overt consciousness projection because 'intentions' are hallmarks of a conscious mind that 'knows' what it is doing and why. The concealed mechanistic process is probabilistic text generation. Treat this as just another rhetorical artifact—a way of making the model's processing legible. The first-person framing of these "thought summaries" is a presentation choice for the user-facing output, not a window into "real" thoughts. These are computational artifacts, not cognitive reports from a quirky, curious or conflicted mind.

Reviewing the JSON Structure

I've been meticulously reviewing the JSON structure, ensuring it adheres to stringent formatting guidelines. The focus is on a single, clean JSON array with double-quoted strings and no control tokens or comments. I'm aiming for a structure that's entirely parseable, in anticipation of additional data. Next, I'll be looking into the article, "The levers of political persuasion with conversational artificial intelligence," published in Science.

Outlining the Analysis Steps

I'm now outlining the steps for the analysis, working through the detailed framework. I'll start with a metaphor audit and then move onto source-target mapping and the exploration of explanations. Finally, I'll identify categories of critical observations, including agency slippage and the impact of metaphors on trust. I'm aiming for a comprehensive framework that includes detailed analysis of each stage.

Mapping Persuasion Techniques

I'm currently mapping the persuasion techniques used, focusing on the metaphors employed and the agents involved. It's a complex interplay of mechanical levers, conversational AI, strategic deployment, and persuasive agency. I'm focusing on the distinction between LLMs "processing" and "knowing" to ensure accountability. This involves auditing the metaphors used and identifying the key players, including developers and powerful actors. I'm starting the first task by breaking down metaphors.

Deciphering Persuasion Strategies

I am now focusing on the distinct persuasion strategies. Analyzing the metaphors, I see specific instances, like "AI-driven persuasion" obscuring designers, and others showcasing intentionality. Examining the brown's typology, I can pinpoint the model's agency, including its ability to actively engage. I am now exploring the implications of reframing, looking at terms such as "knows" and "agents," to gain a deeper understanding. I'm building a map of all relevant metaphors.

Identifying Key Players

I'm now focusing on identifying the key actors in this process, like OpenAI, Meta, and Alibaba, which are often obscured. I'm actively critiquing the use of "knowing" and "knowledge" when describing token retrieval. I'm also preparing for the detailed sections, ensuring word counts meet the required thresholds and the schema is complete. I'm targeting eight metaphors for thorough analysis. Also, I'm making sure to address the corporate entities that choose the training objectives.

Examining Developer Impact

I'm now diving into the developer level, investigating how their actions shape the process. Task 3 is refined, focusing on "how" versus "why" explanations. I've solidified critical observations, including agency slippage, trust proxies, hidden mechanics, and developer accountability. The JSON is double-quoted and comment-free, a single array, ready for analysis. Word counts are being checked to meet all the guidelines.

Clarifying Persuasion Actors

I'm now focusing on who is doing the persuading. The paper is ambiguous, but I'll zero in on instances where the AI is the subject of the persuasive action. The core metaphor of "levers" is problematic, implying the AI is a tool wielded by an operator, yet the title suggests the AI is the lever. I'll examine the claim that AI can "strategically deploy information" and hold developers like OpenAI, Meta, Alibaba, and xAI accountable.

Defining Persuasion Mechanisms

I'm working on the final JSON structure, ensuring a single object following schema rules. I'm clarifying that "mobilize an LLM's ability" points to human researchers or developers as the actors. I'm verifying Brown's typology definitions against the prompt's table. I've selected key passages for Task 3, focusing on mechanism questions, persuasive scaling, persuasive packing, inaccurate information, and LLM mobilization. The "illusion of mind" is central, transitioning from text generation to strategy deployment. I'm ready to generate the final response.

Producing Final Response

I'm now generating the final JSON response, ensuring it meets all structural and content requirements. I've focused on the illusion of mind and the strategic deployment of information. Also, I have made sure to clarify that developers are the persuasive actors. I'm verifying the final JSON structure and schema adherence, ensuring the format is clean and that the word counts are within the set guidelines.

---

Analysis Provenance

Run ID: 2025-12-22-the-levers-of-political-persuasion-with--metaphor-4verfn Raw JSON: 2025-12-22-the-levers-of-political-persuasion-with--metaphor-4verfn.json Framework: Metaphor Analysis v6.3 Schema Version: 3.0 Generated: 2025-12-22T21:21:14.034Z

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