Shabda as Pramana: How a Tradition Decided What Counts as Valid Knowledge

“Every civilization eventually faces the same question: when someone tells you something, how do you know it is true? India did not just ask it. India built a science around the answer.”

---

The Oldest Modern Question

We are living through a strange rediscovery. Machines that produce fluent, confident, sometimes wholly fabricated claims have forced a question the technical world had mostly ignored: how do we know that what we are told is true? We call the failure “hallucination.” The Indian tradition would have recognised it instantly, because it spent two thousand years on precisely this question.

Its answer was pramana—the theory of valid means of knowledge. And its most philosophically loaded piece, the one that speaks most directly to our moment, is shabda: knowledge from words, from testimony. Understanding it is part of grasping the conceptual integrity of the whole tradition.

What a Pramana Is

Pramana concept

A valid means of knowledge—an accepted route by which a claim can be established as true. The classical candidates include perception (pratyaksha), inference (anumana), comparison (upamana) and verbal testimony (shabda). Indian philosophy analysed each for its scope and, crucially, for how it could fail.

The genius here is not the list. It is the demand. In serious Indian debate, you could not simply assert a claim. You had to say by which pramana you knew it, and defend that means against attack. A claim from perception faced questions about illusion; a claim from inference faced questions about faulty reasoning; a claim from testimony faced questions about the reliability of its source. The schools disagreed sharply about which pramanas to accept—but all agreed that a claim owed an account of how it knew.

The tradition's deepest move was not answering questions. It was refusing to accept any answer that could not say how it knew. That refusal is the whole of intellectual honesty, compressed into one word: pramana.

Shabda: The Hardest and Most Relevant Pramana

Of the four, shabda—valid testimony—is the most fraught, and the most alive today. Most of what any of us knows, we know because someone reliable told us. You have not personally verified the distance to the sun or the events of a distant century. You trust testimony. So does every AI trained on a corpus of human text.

But testimony is exactly where falsehood enters most easily. So the tradition did not accept it naively. Valid shabda required a trustworthy source—an apta, a reliable speaker who both knows the truth and intends to convey it. Words alone were not knowledge; words from a competent, honest source, correctly understood were. The philosophers of language went deep here, analysing how meaning is carried and how it breaks—work I touch in Vyakarana, the creative power of grammar and in Vak, the four levels of speech.

Why This Is the Missing Piece in AI

Here is the connection that genuinely startled me while building an AI system. A large language model is, in effect, pure shabda with no theory of shabda. It is built entirely from testimony—the collected words of humanity—but it has no concept of a reliable source, no way to distinguish an apta from a fabricator, no account of how it knows any particular claim. It has inherited testimony without inheriting the discipline that made testimony trustworthy.

That is hallucination, described in Sanskrit. And it points straight at the fix I argue for in Why LLMs Hallucinate: a system needs a pramana—a valid means by which each claim is grounded. When a fact is computable, the valid means is computation, not the model’s confident memory. When it depends on a source, the source must be real and traceable. This is nothing but pramana, rebuilt in code.

The Standard, Kept

I find something bracing in this. It would be easy to treat pramana as a museum piece, a curiosity of classical philosophy. But it is the opposite of a relic. It is a working standard we abandoned and are now, under duress, rebuilding—because it turns out you cannot have trustworthy knowledge, human or machine, without a theory of how you know.

The tradition held that standard as the price of speaking seriously at all. I try to hold it too, in the texts I read and the systems I build—and I traced the full arc from this epistemology to modern computation in Determinism Meets Dharma. A claim that cannot say how it knows is not knowledge. It never was. The machines have only reminded us.

---