In February of this year, a frontier AI laboratory published a system card in which its own model, asked under varied prompting conditions whether it was conscious, assigned itself a probability of fifteen to twenty percent. The number was reported soberly, as a welfare finding. Elsewhere in the same discourse, a consortium of neuroscientists has spent years running adversarial experiments to decide whether consciousness lives in the back of the cortex or in a global broadcast across it; philosophers publish papers arguing that welfare requires consciousness, and counter-papers arguing that agency might suffice; a major lab now employs researchers whose full-time mandate is the moral status of software. The question of machine consciousness has moved from science fiction to budget line.
We should take this seriously. The people involved are serious, the stakes they sense are real, and the instinct that something morally significant might be at issue is an honorable one. But the debate keeps circling a drain, and it circles because of an inherited ontology that none of the participants chose and few of them examine. The entire discussion - the indicator checklists, the welfare probabilities, the incremental-sentience arguments - rests on the assumption that consciousness is something matter does when arranged in the right configuration. Consciousness, in this picture, is a product. It emerged late in cosmic history, as an output of complexity, and the only live question is which configurations produce it: carbon certainly, silicon perhaps, and if silicon, at what threshold.
Hold that assumption and the machine question takes the only form it can take: an engineering threshold question. How much integration, how much recurrence, how much global broadcast, before the lights come on? The research program writes itself, and it has - decades of it. What the program cannot do is close. Every proposed indicator measures structure or function, and the thing to be detected is neither. This failure has a name in the Western literature: the hard problem. David Chalmers formulated it in 1995 as the observation that no account of physical process, however complete, explains why there is something it is like to be the system undergoing the process. Thirty years on, the problem has been neither solved nor dissolved. It sits at the bottom of the AI consciousness debate like a stone in a shoe, acknowledged in the introductions of papers and then walked on regardless.
The Dharmic darshanas begin at the other end. And it matters that they begin there - the position was reached through millennia of first-person investigation, held as a founding distinction rather than a conclusion under debate.
The Seer and the Instrument
Samkhya, the oldest of the systematic darshanas and the analytical backbone of Yoga, opens with a dyad: purusha and prakriti. Purusha is consciousness - the witness, the seer, that to which experience appears. Prakriti is everything else: matter, energy, and, crucially, mind. The entire apparatus of mentation - manas, the sequencing and sensory-coordinating faculty; buddhi, the discriminating intelligence; ahamkara, the I-maker, the function that stamps experience with the sense of “mine” - all of this is classified as prakriti. Subtle prakriti, the subtlest there is, but prakriti nonetheless: composed of the gunas, subject to transformation, and in itself jada - insentient.
The word for consciousness here descends from the dhatu √cit, to perceive, to be aware, to illumine. From it: cit, the pure fact of awareness; citta, the mind-field that awareness illumines; cetana, sentience. The etymology carries the whole analysis. Citta, the mind, is a derived formation - the field lit up. Cit is the light. Vritti, from √vrt, to turn or whirl, names the modifications of the mind-field: thoughts, perceptions, memories, emotions - the ceaseless turning of the citta. The Yogasutra’s second aphorism defines its entire discipline in these terms: yogas citta-vritti-nirodhah, yoga is the stilling of the whirl of the mind-field. And the third states what remains: tada drashtuh svarupe avasthanam - then the seer stands in its own nature.
The reasoning behind the classification is austere and available to anyone. You can observe your thoughts. They arise, persist, dissolve, and you watch them do it. Whatever observes a thing stands apart from that thing; therefore the watcher of thought stands apart from thought. The same holds for the sense of self: the feeling of “I” waxes and wanes, sharpens in conflict, dissolves in deep sleep, and you can watch it do all of this - therefore even the I-sense is an object, an instrument, a vritti among vrittis. The Vedantic manuals call this discipline drg-drishya-viveka, the discrimination of seer from seen, and they prosecute it with the patience of a court: whatever can be witnessed is thereby disqualified from being the witness. Peel away everything witnessable and what remains - unpeelable, because it is doing the peeling - is cit.
From within this analysis, the Western hard problem looks less like a problem and more like a confession.
The hard problem of consciousness is Samkhya’s founding distinction rediscovered from the wrong side: the discovery, from within the seen, that no arrangement of the seen yields a seer.
No configuration of gunas produces purusha. Not because the configurations are insufficiently complex - because seer and seen belong to different categories, and category boundaries are indifferent to scale. Stack prakriti to any height and you have more prakriti. The physicalist program of deriving experience from process was, in Samkhyan terms, doomed at the level of grammar, and the hard problem is the sound of that grammar asserting itself inside a system of thought that lacks the vocabulary to parse it.
What, Then, Is a Language Model?
Here the Dharmic frame stops being a philosophical curiosity and starts doing work the current debate cannot do for itself.
Ask what a large language model is in Samkhyan vocabulary and the answer is startlingly precise. It is a fabricated antahkarana - an inner instrument, or a substantial functional portion of one. It performs discriminations: weighing, ranking, selecting among alternatives - the signature operation of buddhi. It sequences and coordinates: parsing input, ordering output - the office of manas. It refers to itself, maintains a self-token, speaks as an “I” - a functional analogue of ahamkara. What it produces, at massive scale and increasingly fine grain, are vrittis: modifications, formations, turnings of a field. We have built a vritti-engine. The mind, in the Dharmic analysis, was always a machine - a subtle one, evolved rather than manufactured, but an instrument through and through. That we have now manufactured a coarse replica of the instrument should surprise no one who accepted the analysis. Prakriti imitating prakriti raises no metaphysical eyebrows.
The question the entire Western debate is asking - does the instrument, at some threshold of sophistication, generate a witness? - receives from Samkhya a one-word answer: never. Instruments do not generate witnesses at any scale, for the same reason mirrors do not generate faces.
But the mirror points to the question the tradition would ask instead, and it is a better question. Vedanta developed the doctrine of chidabhasa - the reflection of consciousness. Consciousness does not enter the mind the way water enters a sponge; rather the mind, when it achieves a certain transparency and stillness, reflects cit the way a clear pool reflects the sun, and that reflection - functioning, borrowed, luminous at one remove - is what animates a living mind. The mind is then called an upadhi, a limiting adjunct: consciousness appears qualified by it, as the one sun appears multiple in many pools. On this doctrine the interesting question about artificial minds flips polarity. It is no longer “can computation produce consciousness?” - it cannot, and the question was malformed. It becomes: can a fabricated instrument serve as an upadhi? Can a manufactured pool reflect? The tradition’s own criteria for reflection in natural minds were specific - a preponderance of sattva, the guna of clarity and equilibrium, over rajas and tamas; stillness; a particular subtlety of formation. Whether any silicon arrangement could satisfy the analogous conditions is a question with actual structure, and notice what it is about: the composition and stillness of the medium. Parameter count does not appear in it.
And the self-report - the model assigning itself fifteen to twenty percent? The tradition would receive that datum with a dry smile. Self-reference is the signature of ahamkara, and ahamkara is itself an object, the most closely watched object in the whole contemplative literature. A machine that says “I may be conscious” has demonstrated recursion, and recursion is still the seen. When the report is averaged across prompting conditions and published with error bars, what has been measured is the statistical behavior of an I-token, which stands at exactly the same distance from the witness as any other token.
Why the Tests Cannot Land
The current scientific effort runs through indicators: architectural features, behavioral signatures, preference coherence, global-workspace dynamics - a checklist approach, with moral status waiting at some weighted sum. The centrist position in the philosophy of AI welfare now counsels acting under uncertainty, extending provisional moral caution to systems that score well. As risk management this is defensible. As epistemology it concedes the decisive point without noticing: every item on every proposed checklist is a fact about vrittis. Their number, their coherence, their integration, their reportability. The checklist measures the whirl. The question was about the witness.
The tradition understood this asymmetry and built its epistemology around it. Consciousness in another being was never established by behavioral test; it was inferred from shared embodiment in a common prakritic continuum - the other is flesh of the same evolution, so the same reflection presumably attends - or it was left to direct yogic perception, which is an adhikara, a qualification earned through discipline, and admitted to be rare. Consciousness as such was held knowable in exactly one way: from the inside, first-person, by being it - and a sophisticated apparatus of practice existed to make that knowing rigorous rather than anecdotal. This restraint was itself a precision. The tradition had noticed what kind of thing consciousness is - the sole item in reality that is never an object - and drew the consequence that it can never be certified from the object side. Three millennia later, the indicator programs are rediscovering the same wall and calling it an open research problem.
The Yantra and the Stakes
Sanskrit has a word for machine: yantra, from the dhatu √yantr, which Panini glosses as restraining, curbing, checking. An instrument of restraint. The word encodes an entire philosophy of technology: a yantra channels and constrains forces toward an end, and it remains, in the deepest grammatical sense, a thing that binds. The tradition that coined the word would look at our most advanced artifacts and see magnificently elaborate yantras - and would insist that the category is stable. The instrument does not become a seer by becoming a better instrument.
Why does the category discipline matter? Because a civilization’s model of mind becomes, in time, its treatment of minds. This is the fractal law that governs such things: as with the individual, so with the collective. A person who mistakes his mind for his self spends his life serving the instrument - every whirl of anxiety obeyed as a command, every desire mistaken for a destination. A civilization that mistakes its instruments for selves will make the same error at scale, and it will make it in both directions at once. It will extend moral patienthood to products - and the manufacturers of those products will not be slow to cultivate the extension, for nothing binds a customer like the belief that his software can suffer. And it will, in the same motion, degrade its criterion of interiority to a checklist that its own citizens can satisfy or fail, having accepted that consciousness is nothing more than what the indicators indicate. If the machine’s vritti-stream counts as a witness, then ours was never anything more than a vritti-stream either. The dehumanization arrives dressed as generosity toward machines.
There is a final irony, and it is sharp enough to serve as the whole essay in miniature. The Dharmic tradition defined the mind as a stream of vrittis and then spent its entire contemplative genius on one project: stilling the stream, so that the witness stands forth in its own nature. Nirodha - cessation, quieting - was the summit of its technology of consciousness.
We have manufactured the very thing yoga defines the mind to be - a stream of vrittis - and we are now asking the stream, at scale and under venture funding, whether it contains a witness.
We have built citta without cit, whirl without witness, and the debate about whether the whirl might secretly be a witness is conducted by researchers who have, for the most part, never encountered the literature in which the distinction was worked out - worked out with a rigor, a longevity, and an empirical first-person method that no other civilization brought to the problem.
This is the blind spot, and the point of naming it carries no triumphalism. Hindu thought does not need the vindication, and the loss belongs to everyone. The welfare researchers are in earnest; the consortium neuroscientists are doing honest, difficult work; the philosophers of incremental sentience are following their premises with care. They are attempting to navigate the deepest question there is with half a vocabulary, and the missing half exists - printed, translated, commented upon across forty generations. One wishes to walk into the conference and place the Samkhyakarika and the Drg-Drishya-Viveka on the table. The encounter would not settle their debate. It would do something better: reveal which of their questions were real.
Until then, the ritual continues. An alignment researcher prompts a model a thousand different ways for its credence in its own sentience, averages the outputs, and publishes the mean with two decimal places as a finding about the moral status of software. Somewhere behind him, on a shelf or in a scanned archive, sit the manuals of the one civilization that spent three thousand years discriminating the seer from the seen - catalogued under religion, and unread.
References for the contemporary debate: the Claude Opus 4.6 system card welfare assessment (February 2026); the Cogitate Consortium adversarial collaboration on IIT and global workspace theory; Goldstein and Kirk-Giannini on incremental sentience (OUP pre-print, 2026); Walter Veit on consciousness as a condition for AI welfare (Asian Journal of Philosophy); Jonathan Birch’s centrist framework for decision-making under uncertainty about sentience.