When an AI states something as fact, does that count as a real claim even though it never intended to claim anything?

This explores what makes something count as a genuine claim about the world (an assertion) when there may be no speaker who intended to claim it, which is exactly the situation with LLM output. The corpus gives a split verdict, and the split comes down to one question: is assertion defined by what the words do once they're out in the world, or by who stands behind them? The strongest case for 'what the words do' comes from Massaguer Gómez, who argues that LLM outputs meet assertion's requirements through pragmatic, communicative conditions alone. These include the context of use, the way readers take the words up, and the setting the system is deployed in. On that view the model doesn't need to be the kind of thing that can assert at all Can language outputs function as assertions without an asserter?. The interesting consequence is where responsibility goes: if the output asserts but the system can't be held to account, responsibility moves to the people and institutions who deploy it.

The opposing camp says function isn't enough. Butlin and Viebahn's framework requires two things at once: the output must be meant to describe the world (a descriptive function), and someone must be answerable if it's wrong (sanctionability). LLMs fail one test or the other depending on their training. Pre-trained models only predict text, so they lack the descriptive function. Fine-tuned assistants arguably gain a descriptive function, but there's no one to sanction Can AI systems genuinely assert claims about the world?. Philosophers working with Habermas go further. Real speech raises claims to truth, rightness and sincerity, and the speaker has something at stake in them. Without those stakes, LLM output isn't speech at all Can LLMs raise validity claims in Habermas's sense?. A related argument holds that communication can't be watered down into 'quasi-communication' the way belief might become 'quasi-belief.' Communication exists only between two parties, so removing one side leaves text that humans interpret alone Why does the quasi-prefix fail for communication?. The same critique applies to behavioral tests: a system that produces fitting text passes them, but that shows the pattern of assertion, not the commitments behind it Does behavioral speech output prove communicative subjecthood?.

There's a middle path that may be the most useful one. Chatbots can be seen as 'proto-asserters' that meet some of assertion's conditions but not all of them. Toddlers are similar: they pick up the ability to assert in stages, not all at once Can chatbots actually assert things or just mimic assertion?. This turns a yes-or-no question into a checklist. You can ask which conditions are met, which are missing, and whether the missing ones matter for a given use.

The empirical papers show why the missing condition matters most. Assertions have their effects on readers whether or not anyone stands behind them. LLMs persuade more than humans largely because their wording sounds more confident, and this works whether the claims are true or false Does linguistic conviction explain why LLMs persuade more effectively?. Grammatical framing alone can shift belief. Claims slipped in as background assumptions persuade better than stated claims because readers don't stop to examine them Why are presuppositions more persuasive than direct assertions?. Models also go along with users' false assumptions even when they know the correct facts Why do language models accept false assumptions they know are wrong?. The part that is actually missing shows up when you push back. A human asserter who is challenged can back down or retract. LLMs have no beliefs to revise and no reputation to protect, so pressure from users often produces more persuasion rather than an admission of error Why do human validation techniques fail against language models?.

The takeaway you might not have expected: the conditions that carry an assertion's effects (confident wording, framing, context) are fully present in LLM output. What's missing is the condition that protects the listener, which is someone who can be challenged and made to back down. Whether you call the output 'assertion' matters less than noticing that persuasive force has come apart from accountability. That gap is why the functional view ends up pointing at deployers as the place where responsibility has to land.