Assertion Without a Speaker: Testimony, Tracking, and Large Language Models

The argument starts from what the leading accounts of testimony, those of Burge, Williamson, Goldberg, Faulkner and Lackey, share beneath their disagreements. I call it the Tracking Minimum: a source can pass on knowledge that p only if the truth of p, or the evidential support for p, figures in the explanation of its producing the assertion that p, and figures there in virtue of the source's own operation rather than of a proxy on which that operation depends. The last clause, the endogeneity requirement, carries most of the weight.

Current systems are often accurate, and increasingly so, but accuracy is not tracking. On the best current description, their open-domain answers are governed by likelihood under patterns learned from human text, shaped by training against human approval, and the truth reaches those answers only through a proxy. The paper weighs the interpretability evidence for truth-tracking internal representations, the case for AI testimony made by Freiman and Faria, and the strongest objections, and shows that the remaining disagreement turns on an empirically decidable question, for which an experiment is specified.

The positive proposal is that relying on a model is instrument reading without transparent calibration: the user's warrant, when present, is generated by her own use and not transmitted to her. The result is less comfortable than a reclassification sounds, since a working barometer satisfies the Tracking Minimum and these systems fail a condition ordinary instruments meet by construction. And because an AI link preserves no one's grounds, chains of AI-generated text reposted by users do not preserve warrant as chains of human testimony do. They preserve text.

The main claims

1. The five leading accounts of testimony all entail the Tracking Minimum.
2. The Tracking Minimum is explanatory and endogenous: the truth must figure in why the source asserts, and in virtue of the source's own operation rather than of a proxy.
3. Even Lackey's permissive statement view, which drops the belief condition, still requires it.
4. On the best current evidence, current language models satisfy it neither originally nor derivatively, so they do not transmit testimonial knowledge.
5. Independently of that empirical question, hearers are not now positioned to accord their outputs testimonial uptake.
6. Relying on a model is instrument reading without transparent calibration; a working barometer meets the condition these systems fail.
7. Chains of AI-generated text do not preserve warrant. They preserve text.

Abstract

Large language models produce assertion-shaped outputs that often bring their users to true belief. I argue for two conclusions of different strengths: that on the best current evidence these systems do not transmit testimonial knowledge, and that hearers are not now positioned to accord their outputs testimonial uptake, which does not wait on the first. From Burge, Williamson, Goldberg, Faulkner, and Lackey I derive a Tracking Minimum: a source transmits testimonial knowledge that p only if the truth of p, or the evidential support for p, figures in the explanation of its producing that assertion, and does so in virtue of the source’s own production process rather than of a proxy on which that process depends. It is this endogeneity requirement, not any modal profile, that the five accounts share and current systems fail.

The argument engages Burge’s extension of his framework to computational sources, the interpretability literature, and the published case for AI testimony, whose dispute with mine turns on an empirically decidable question, for which an experiment is specified. The positive proposal places LLM outputs within an extended epistemology of instruments, with a result that has gone unremarked: a working barometer satisfies the condition these systems fail.

@article{saxena2026assertion,
  author  = {Saxena, Abhinav},
  title   = {Assertion Without a Speaker: Testimony, Tracking, and Large Language Models},
  journal = {Episteme},
  year    = {2026},
  pages   = {1–22},
  doi     = {10.1017/epi.2026.10160},
}