Nearly two millennia ago, Indian thinkers were preoccupied with a remarkably similar difficulty: what makes a claim worthy of belief?
The Nyāya tradition developed an elaborate theory of pramāṇa, or valid means of knowledge. Its classical account recognised perception, inference, comparison and trustworthy testimony.
India could make these traditions part of serious interdisciplinary research involving philosophers, computer scientists, educators and public administrators.
A chatbot can answer a difficult question in five seconds. Ask it why its answer is correct and it may produce another five paragraphs. The difficulty is that the explanation can be as unreliable as the original answer.
We have built machines that speak with authority. We are still learning how to make them deserve it.
Nearly two millennia ago, Indian thinkers were preoccupied with a remarkably similar difficulty: what makes a claim worthy of belief?
The Nyāya tradition developed an elaborate theory of pramāṇa, or valid means of knowledge. Its classical account recognised perception, inference, comparison and trustworthy testimony. These were not interchangeable routes to truth. Each had its own strengths, conditions and possibilities of error.
Consider inference. Smoke may suggest fire, but only if the relationship is properly established and alternative explanations have been considered. A plausible connection is not automatically proof. The Naiyāyikas even catalogued the ways reasoning fails, calling them hetvābhāsa: reasons that only look like reasons. Mist rising from a hill, they warned, can be mistaken for smoke. Many artificial intelligence (AI) errors are simply mist dressed up as smoke.
Advaita thinkers offered a sharper image: a glint of mother-of-pearl on the shore mistaken for silver. The silver is not there, yet the mind sees it vividly. That is about as good a description as any of what technologists now call hallucination.
That distinction is painfully relevant to modern AI.
A language model predicts and generates responses from learned patterns. Its fluency does not guarantee factual accuracy. A convincing citation may be fictitious. An apparently logical conclusion may rest on a mistaken premise.
Sharpening The Questions
Nyāya cannot be transplanted into computer code as a ready-made solution. It can, however, sharpen the questions we ask when evaluating AI: What is the evidence? How was the conclusion reached? Which assumption could be wrong?
Nyāya demanded vyāpti, an invariable connection, not a frequent one. The link between smoke and fire was checked in the kitchen hearth, where both appear, and against the lake, where neither does. A language model learns what usually appears together. Nyāya asked what must.
Nyāya accepted testimony only from an āpta, someone who knows the matter first-hand and has no motive to mislead. A chatbot has read about the village. It has never walked through it.
There is another Indian intellectual practice worth recovering: pūrvapakṣa, the careful presentation of a position before examining or challenging it. In its strongest form, it demands that we understand an argument before attempting to defeat it.
Imagine applying that discipline to an AI system advising a policymaker. Instead of presenting only its preferred recommendation, the system would also explain the strongest objections and identify who might lose.
In district administration, this is not an academic luxury.
A proposal to construct a road may promise shorter journeys and better market access. The revenue records may show government land. Yet a field visit may reveal that villagers use a part of that land for a weekly market or a traditional community activity.
The file contains facts. The village contains other facts.
Data: Not False, Merely Incomplete
As the district collector of Latehar, Jharkhand, I once saw a dashboard report near-total toilet coverage in Alaudia panchayat of Chandwa block in 2006. In its villages, several of those toilets had quietly become storerooms for firewood and grain. We shifted our focus from construction to behavioural change, persuading families to actually use what had been built. Looking back, the data was not false, merely incomplete. And incompleteness delivered with confidence is the most dangerous error of all.
In 2017, as collector of Khunti, Jharkhand, I saw satellite images of a crop-loss assessment showing fields lush and green. When I walked through the villages of Torpa block, the green turned out to be weeds that had overrun failed paddy.
Deep in the Saranda forest of West Singhbhum, during a visit as divisional commissioner in 2020, I met an elderly widow who had been denied rations because the machine could not read her worn fingerprints. It took human intervention, not a software update, to put food on her table.
A system can be accurate about the data and still be wrong about the person.
A responsible decision requires the file and the village.
The Indian knowledge system is also notable for its disagreements. Within it, Buddhist philosophers, Jain thinkers and the Vedic schools contested fundamental questions about truth, reality and reasoning. Their methods were not identical; their arguments were often fierce. Nyāya even separated vāda―debate in pursuit of truth―from jalpa, debate merely to win. Much of today’s AI is tuned for jalpa: answers built to please rather than to be right.
That is precisely why this history matters. It offers a record of inquiry, not a museum of unanimous answers.
As governments and companies worldwide debate AI reliability, India could make these traditions part of serious interdisciplinary research involving philosophers, computer scientists, educators and public administrators.
The objective would not be to claim that ancient India invented AI. It would be to contribute something modern AI badly needs: disciplined doubt. Nyāya itself counted saṃśaya, doubt, among the starting points of all inquiry.
The next great Indian export may be a way of thinking that refuses to confuse a confident answer with a correct one.
(The author is an IAS officer and the author of eight books on artificial intelligence)
(Views expressed are personal)
