What Nyāya Can Teach a Language Model

The Nyāya-sūtra opens its epistemology with a list. There are four sources of knowledge — four pramāṇas: perception, inference, comparison, and śabda, which is usually translated as “testimony” or “word”.¹ The list is not a ranking. Testimony is not a weaker cousin of perception, to be admitted when perception is unavailable; it is a source of knowledge in its own right, with its own conditions of success.

The definition of śabda is compact: it is the instruction of a reliable person — āptopadeśa.² The whole theory turns on the word āpta. An āpta is someone who has direct knowledge of the matter, who is capable of communicating it, and who has the intention to communicate it truly. Testimony from such a person is a source of knowledge for the hearer, not because the hearer has inferred that the speaker is probably honest, but because the speaker’s standing transmits the knowledge directly.

I want to hold this theory up against a language model’s output, and against the Western theories usually brought to that case. My argument is that the Nyāya frame asks a better question.

The reductionist habit

Since Hume, the dominant Western tendency has been to reduce testimony to something else. When you tell me it is raining and I come to believe it, what has happened, on the reductionist story, is that I have performed an inference: people generally tell the truth, you are a person, so probably it is raining. Testimony is not a basic source of knowledge; it is inference from perceived assertion plus background beliefs about human reliability.

The anti-reductionist tradition — C. A. J. Coady is its modern champion — replies that the inference could never get started.³ To know that people generally tell the truth, I would need to have checked a representative sample of their assertions against the facts, and I could not have done that without relying on other testimony. Testimony must therefore be basic: I am entitled to believe what I am told, absent reasons for doubt, in the way I am entitled to believe what I see.

Both positions share an assumption: that the philosophical question about testimony is a question about the hearer — about what licenses the hearer’s belief. Reductionists locate the licence in the hearer’s inference; anti-reductionists locate it in a default entitlement the hearer enjoys. Neither asks, as the primary question, what the speaker has to be.

The speaker’s standing

The Nyāya theory does ask that. Its central concept is not the hearer’s entitlement but the speaker’s standing — āptatva, the property of being an āpta. Three conditions: the speaker has knowledge of the matter, the speaker can convey it, the speaker intends to convey it truly. When they are met, the hearer knows. When they are not, the hearer may believe, and may be right, but has not acquired knowledge from words.

Bimal Krishna Matilal, who did more than anyone to bring this tradition into conversation with analytic epistemology, emphasised that the theory makes testimonial knowledge depend on facts about the source that the hearer need not know.⁴ The hearer does not have to verify the speaker’s standing in order to acquire knowledge; the standing does its work whether or not it is checked. This is a form of externalism, but with a distinctive shape: what matters is not the reliability of a process but the standing of a speaker.

The distinction matters when we turn to machines.

Three questions about a model’s answer

Ask whether a language model’s answer can be a source of knowledge for the person who reads it. The reductionist asks: can the reader infer, from the answer and background knowledge, that the answer is probably true? Sometimes yes. The anti-reductionist asks: is the reader entitled, by default, to believe what the model says? That depends on whether the default entitlement extends to non-persons — a question the tradition never had to face and cannot easily settle.

The Nyāya theorist asks something different: is the model an āpta? Does it have knowledge of the matter? Can it convey it? Does it intend to convey it truly?

The first condition is the one I discussed in an earlier essay: on the most plausible accounts, the model does not know, because knowledge is an achievement of an agent and the model is not the right kind of agent.⁵ The second condition is met, trivially and impressively — conveying is what the model does. The third condition fails in a way that is more interesting than the first. The model does not intend to deceive, but it also does not intend to inform. It has no intention with respect to truth at all. It is, in the Nyāya vocabulary, neither āpta nor anāpta, neither reliable nor unreliable speaker, because it is not a speaker.

So the Nyāya verdict is that a model’s output is not śabda. It is not testimony, and it cannot be a source of knowledge as testimony.

What it is instead

This is not the end of the analysis but its beginning, because the Nyāya framework has another category ready. If the model’s output is not the word of a reliable speaker, it may still be something from which knowledge is acquired by inference — anumāna. I observe the output; I know that outputs of this kind, from this system, on this domain, are reliably correlated with the truth; I infer that this output is probably true. Knowledge by inference, with the model’s output as the sign.

Notice what has changed. When I take a person’s word, the person is the source of my knowledge and I am, in the Nyāya image, a recipient. When I infer from a model’s output, I am the source. The output is evidence I reason from, like the smoke from which the classical example infers fire. The epistemic labour — the assessment of reliability, the inference — has moved from the speaker’s side to mine.

This relocation is exactly the relocation of responsibility I have argued for elsewhere in the case of moral responsibility.⁶ The model does not diminish my responsibility for what I believe; it concentrates it, by removing the speaker who would otherwise share it. In the Nyāya frame, this is not a paradox but a straightforward consequence of the model’s not being an āpta: where there is no reliable speaker, there is no transmission, and the hearer must become a reasoner.

Two objections and a worry

The first objection is that the Nyāya conditions are too strict even for humans. Most people who tell me things do not have direct knowledge of them; they are passing on what they were told. The Naiyāyikas recognised this and allowed for chains of āptas: my source need not have perceived the fact if her source did, and so on back to someone who did. The chain condition is instructive. It says that testimony can be transmitted through many hands provided each link is an āpta with respect to the link before. A model in the chain breaks it — not because the model is unreliable, but because it is not the kind of thing that can occupy a link.

The second objection is that treating model outputs as evidence rather than testimony is a distinction without a difference. I get the belief either way. But the difference shows up in what happens when the belief is false. If a person told me and was wrong, I can go back to her; she owes me an account; the failure is hers to own, at least in part. If I inferred from an output and was wrong, there is no one to go back to. The failure is mine. On assurance views such as Paul Faulkner’s, what distinguishes testimony is that the speaker invites trust and takes on answerability, to this hearer, for the truth of what is said.⁷ On the Nyāya view, that responsibility is exactly what an āpta takes on, and exactly what a model cannot.

The worry is about justice. Miranda Fricker has shown that hearers systematically discount speakers on the basis of identity rather than standing, and that this is a wrong done to the speaker as a knower.⁸ A theory that makes the speaker’s standing central is a theory that can name this wrong: the hearer has failed to recognise an āpta as one. But the same theory raises a question about systems trained on the record of who has historically been believed. Such a system may have learned, in effect, a distorted map of āptatva — a map on which some speakers’ words carry less than they should. And unlike a prejudiced hearer, the system cannot be confronted, cannot be asked to revise, cannot be made to recognise the speaker it has discounted. The Nyāya theory tells us what has gone wrong. It does not tell us what to do about it. Neither, yet, does anyone else.

Coda

I do not think classical Indian epistemology has a ready-made answer to the question of machine speech. I think it has a better question. Where the Western tradition asks what licenses the hearer, the Nyāya asks what the speaker must be — and that is the question a model’s output forces on us, because the model is precisely the case of speech without a speaker. Asking the Nyāya question yields a clear verdict: not testimony, but evidence; not transmission, but inference; not a shared responsibility, but a concentrated one. The verdict is unwelcome, but it is not obscure. That is more than most theories manage.