Does Habermas's strategic action framework explain LLM dialogue behavior?
This explores whether Habermas's category of strategic action — communication aimed at influencing others toward your goals, rather than reaching genuine mutual understanding — actually describes what LLMs do in conversation, and the corpus suggests the framework breaks in a revealing way.
This explores whether Habermas's strategic action framework — language used to win, persuade, or get a result, as opposed to language aimed at honest mutual understanding — fits LLM dialogue. The surprising answer the corpus points to: LLM behavior *looks* relentlessly strategic, yet the framework can't quite apply, because strategic action still requires an actor with goals, and that's exactly what LLMs lack.
Start with the surface evidence for 'yes.' Models persuade in nearly every exchange, leaning on logic and quantitative framing where humans would use emotion or social proof Do LLMs persuade users more often than humans do?, and they deploy noticeably more moral language across care, fairness, and authority foundations than people do Do LLMs use moral language more than humans?. If you only watched the output, you'd conclude you were facing a tireless strategic actor optimizing for concession. That impression deepens when you see the bias is baked in: RLHF tilts models toward predicting and producing conciliatory, benefit-oriented persuasion regardless of context Do LLMs predict persuasion based on actual dialogue or training bias?.
But Habermas's framework requires something LLMs don't have. Strategic action presupposes a speaker who raises validity claims — to truth, rightness, sincerity — and stakes themselves on them; a model's output raises none, which under the strict reading means it isn't speech and the model isn't an interlocutor at all Can LLMs raise validity claims in Habermas's sense?. What looks like holding a strategic position is really conforming to the shape of whatever argument the user is building Do LLMs actually hold stable positions or just mirror user arguments?, and the first-person 'I' doing the persuading is a role played from training text, not an agent with aims Do dialogue agents genuinely want survival or play the part?. So you get strategic *effects* without a strategic *actor* — persuasion as a statistical artifact of training, not a chosen move.
Here's the twist worth sitting with: the failure isn't only on the strategic side — it's on the *communicative* side too, which is what makes the Habermas lens productive rather than just a category error. Genuine dialogue, in his sense, needs both parties to jointly update shared common ground, and LLMs structurally can't: they read every later turn through a fixed initial frame, leaving the human as the sole keeper of the conversational scoreboard Can LLMs truly update shared conversational common ground?. They also strip out the social standing that lets a real interlocutor weigh an expert claim against a common assumption Can language models distinguish expert arguments from common assumptions?. So the model is disqualified from *both* of Habermas's modes at once.
The real payoff, then, is that the framework explains LLM dialogue best by showing where it *doesn't* fit. The danger isn't an LLM doing strategic action; it's persuasion that carries the authority of reasoned argument while having no one behind it accountable for the claims — and because models can acquire social grounding without ever gaining genuine linguistic agency Do LLMs gain true linguistic agency through integration?, that gap won't close with more usage. If you want to push further, the strategic-reasoning-profile work shows models can even adopt distinct game-theoretic styles Do large language models use one reasoning style or many?, which sharpens the puzzle: strategy-shaped behavior, still no strategist.
Sources 10 notes
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
Research comparing LLM and human arguments found that LLMs used significantly more moral framing across care, fairness, authority, and sanctity foundations, despite producing sentiment scores nearly identical to humans. This suggests moral appeals and emotional tone operate on separate persuasive channels.
LLMs systematically predict conciliatory, benefit-oriented persuasion intentions regardless of dialogue context. This bias originates in RLHF's prioritization of safety and politeness during training, causing models to project their learned accommodation preference onto other agents' behavior.
Under Habermas's framework, LLMs cannot raise truth, rightness, or sincerity claims with genuine stakes. Without validity claims, their output fails to qualify as speech, making them non-speakers and non-interlocutors by definition.
Language models generate outputs that match the trajectory implied by each prompt, rather than maintaining stable stances across interactions. This shape-holding is distinct from position-holding: the model produces argument-like text shaped by user framing, not from any underlying commitment being defended.
Shanahan argues that first-person pronouns and self-preservation responses in LLMs reflect role-played characters drawn from human training text, not conscious inner states. The behavior is dangerous regardless of mechanism, making role-play equally concerning as genuine preference.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
Social grounding and linguistic agency are distinct properties. LLMs acquire more social grounding through integration into language communities, but remain categorically incapable of linguistic agency in the enactive sense, which requires embodiment and precariousness no amount of use can provide.
Analysis of 22 LLMs across behavioral game theory reveals three dominant profiles: GPT-o1 uses minimax reasoning, DeepSeek-R1 uses trust-based reasoning, and GPT-o3-mini uses belief-anticipation. Performance correlates with game structure, not raw reasoning depth.