Can AI Testify? Language Models and Testimony

Can AI testify, or is a language model's answer something else?

Short answer

Not in the sense that matters for knowledge. On the best current evidence, large language models do not transmit testimonial knowledge: the truth of what they say does not figure, through their own operation, in why they say it. Their answers are often true and often useful, but relying on them is closer to reading an instrument than to taking someone’s word.

By Abhinav Saxena· October 2026· 3 min read

A patient asks a chatbot what dose of acetaminophen is safe for her ten-year-old. The chatbot answers correctly. She gives the dose, and the child is fine. She came to a true belief and acted well on it. Did she learn the dose from the chatbot in the way she would have learned it from her paediatrician? If the question is whether she came to know the safe dose, as one knows what a competent informant has told one, the answer is less obvious than the case looks.

What testimony requires

When you learn something from another person’s word, you come to know it on her say-so. Philosophers disagree about what makes that possible. Burge grounds it in an entitlement to accept what rational sources present as true; Williamson, in the knowledge norm that governs assertion; Goldberg, in the reliability of the speaker’s own production; Faulkner, in the assurance a speaker gives and the answerability she takes on; Lackey, who has done most to weaken the conditions on speakers, in a statement that reliably indicates the truth.

Read carefully, these accounts converge on one minimal condition, which I call the Tracking Minimum. A source can pass on knowledge that p only if the truth of p, or the evidence for it, figures in the explanation of why the source asserted p, and figures there through the source’s own operation rather than through some proxy on which that operation depends.

Why accuracy is not the issue

The argument is not that language models are unreliable. In many domains they are highly reliable, and successive systems have become more so. It is that the reliability they exhibit is not the kind testimony requires.

There are two species of reliability. One is a high rate of true outputs under the conditions a system was trained and tested on. The other is counterfactual: vary the truth, and the output varies with it. A thermometer that is broken but sits in a room held at a constant temperature has the first kind perfectly and the second not at all. Nobody thinks that posting its readings passes on knowledge of the temperature.

On the best current evidence, what governs the open-domain answers of a large language model is likelihood under the patterns it has learned from human text, shaped by training against human approval. The truth reaches those answers only through a proxy: the fact that what is true tends to be what gets written down. Hold the proxy fixed and vary the truth, and the answer does not move. That is why accuracy, however high, does not settle the question.

What relying on a model is instead

Users plainly come to true and often justified beliefs through these systems, so the negative verdict needs a positive successor. My proposal is that the user stands to a model as a competent reader stands to an instrument. No warrant is passed to her from the system. Her warrant, when she has it, is generated by her own use: by the system’s record in conditions like hers, her competence in reading it, and the absence of reasons for doubt. I call this instrument reading without transparent calibration.

The proposal is less comfortable than it sounds. A working barometer meets the condition these systems fail: vary the pressure and the reading moves, in virtue of the instrument’s own mechanism. So the trouble is not merely that users cannot see how a model is calibrated. It is that the model is not, in respect of the particular content it emits, calibrated to anything.

Two conclusions, held at different strengths

The first conclusion, that current systems do not satisfy the Tracking Minimum, is an empirical claim about what governs their production, held with the confidence the evidence supports and no more. If interpretability research shows that a system’s answers are governed by representations that track the truth, that system meets the condition and the verdict changes for it. That is not a concession extracted under pressure; it is what a structural constraint is for.

The second conclusion does not wait on the first. A hearer who knows what these systems are has standing grounds for doubt, and those grounds suspend the default entitlement to take a source at its word. Whatever the final answer about their inner workings, we are not now in a position to treat their answers as testimony.

Questions people also ask

Does this mean the answers of AI systems are worthless?

They are not worthless, since instruments are genuine sources of knowledge; a well-calibrated thermometer lets you know the temperature. The claim is about the structure of the warrant, not its value. When you come to know something through a model, your warrant is generated by your own use of it, and the work of checking that it is reliable in the case at hand falls on you.

Isn't human testimony unreliable too?

Often it is, but the question is not how often a source is right. A person who tells you what she knows is answerable for it, and the truth of what she says figures in why she says it. A source can be right most of the time without that being so, as a broken thermometer in a room held at a constant temperature is right every time.

Is the argument that machines cannot think or understand?

The argument does not turn on the source being human, conscious or capable of intention, and I do not deny that these systems may understand what they produce. The condition at issue is third-personal and could in principle be met by a machine. It is a claim about what governs the production of current systems’ answers.

What about systems that search the web and cite sources?

Retrieval is real, and it matters; the open question is whether generation is governed by the evidential weight of what is retrieved, or by its presentation, such as coherence, specificity and how authoritative a passage looks. That question can be tested, and until it is answered for a given system, citations show what the answer was conditioned on, not that its truth explains why it was given.

Sources and further reading

Saxena, A. (2026). Assertion without a speaker: Testimony, tracking, and large language models. Episteme, 1–22.doi.org/10.1017/epi.2026.10160
Burge, T. (1993). Content preservation. The Philosophical Review 102(4): 457–488.doi.org/10.2307/2185680
Lackey, J. (2008). Learning from Words: Testimony as a Source of Knowledge. Oxford: Oxford University Press.
Freiman, O. (2024). AI-testimony, conversational AIs and our anthropocentric theory of testimony. Social Epistemology 38(4): 476–490.doi.org/10.1080/02691728.2024.2316622
Faria, D. (2025). Can conversational AIs testify? Challenging the conservative view. Logos and Episteme 16(3): 249–261.