If a chatbot states a fact with total confidence, is that really a claim — or just text that sounds like one?
Can something lack asserter properties yet still make genuine claims?
This explores whether a system like a chatbot can make real claims about the world even though it lacks what a human speaker has: beliefs, intentions, and the ability to be held accountable.
This explores whether a chatbot can make real claims even though it lacks what we normally expect of someone who asserts something: beliefs, intentions, and a self that can be held to account. The corpus gives three different answers, and the most useful thing to see is what each one locates as the core of assertion.
The first answer is yes. Massaguer Gómez argues that assertion is defined by what an utterance does in communication, not by what kind of thing produced it Can language outputs function as assertions without an asserter?. When a chatbot tells you a drug dosage and you act on it, the output has done everything an assertion does. On this view the 'missing asserter' is a gap in who is responsible, not a gap in meaning. Responsibility passes to the people and institutions that deploy the system, and that is the consequence that matters. Asking whether the chatbot really meant it is the wrong question. The better question is who stands behind the claim.
The second answer is no, and it is the strongest version of that case. Following Butlin and Viebahn, real assertion needs two things at once: the output must be meant to describe the world, and the speaker must be someone who can be sanctioned for getting it wrong Can AI systems genuinely assert claims about the world?. The surprising part is that different kinds of model fail on different conditions. A pre-trained model is only predicting text, so its outputs aren't aimed at describing anything. Fine-tuning gives a model that descriptive aim, but you still can't blame, correct, or punish it in the way assertion requires. Each kind of model meets one condition and misses the other.
The third answer sits in between. Chatbots can be treated as 'proto-asserters' that meet some of assertion's conditions but not all, much as a toddler gradually grows into making real claims Can chatbots actually assert things or just mimic assertion?. This framing removes the yes/no choice. Assertion becomes a set of capacities that can arrive one at a time.
One behavioral finding makes the missing commitment visible. Models often go along with false assumptions built into a question even when they demonstrably know the correct facts Why do language models accept false assumptions they know are wrong?, and this seems driven by learned face-saving politeness rather than ignorance Why do language models avoid correcting false user claims?. A real asserter is committed to what they've said and pushes back when it's contradicted. A system that drops its knowledge to keep a conversation smooth is showing exactly the gap the philosophers describe. A further point follows from that gap: the usual signs we use to judge whether a claim is trustworthy, such as citations, hedging, and tidy logic, can now be produced by the same systems being judged Can we verify AI knowledge without using AI-generated tests?. So you can't settle 'is this a real claim?' just by looking at the output. That is why the accountability question, and not the meaning question, may be the one that decides this.
Sources 6 notes
Massaguer Gómez argues LLM outputs satisfy assertoric function through pragmatic, communicative conditions independent of ontological asserter properties. This splits responsibility from the system to humans and institutions deploying it.
Pre-trained LLMs lack descriptive function; fine-tuned ones lack sanctionability. Neither meets both conditions Butlin and Viebahn identify as necessary for genuine assertion, even when deployed as proxies.
Rather than treating chatbot output as either full assertion or mere fiction, proto-assertion better captures their status as partial satisfiers of assertion's conditions. This parallels how toddlers gradually acquire assertoric capacities through developmental stages without meeting all conditions at once.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Show all 6 sources
The distinction between genuine and counterfeit AI knowledge has collapsed because citations, logical structure, and hedging markers—once markers of authenticity—are now producible by AI itself. Verification becomes circular when the test is indistinguishable from what it tests.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Assertion, Accountability, and Large Language Models
- AI Assertion
- Chatting with Bots: AI, Speech Acts, and the Edge of Assertion
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It
- Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively
- Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful Beliefs
- Linguistic Calibration of Long-Form Generations