The debate over machine consciousness is usually framed as a contest between two camps. Functionalists say consciousness is whatever the right kind of information processing does, so a sufficiently sophisticated AI will be conscious. Biological naturalists say consciousness depends on something specific to living brains, so it won’t.

Both camps share an assumption they rarely examine: that consciousness is the kind of thing that gets produced, and that the only question is what can produce it.

The contemplative traditions start somewhere else. They begin with what consciousness is actually like when it is examined directly, from the inside, by a mind trained to look. What they report changes the question about AI. It also exposes a weak argument that is often made in the tradition’s name, and points to a stronger one.

Knowing Without Concepts

From the contemplative perspective, the nature of mind is non-conceptual knowing. Thoughts, labels, categories and stories arise within awareness, but none of them is its nature. Before a sound is recognized as a bird, there is the hearing. Before a feeling is named as sadness, it is known. And beneath every moment of perception and thought there is the plain cognizance in which it all appears, which the Tibetan Buddhist Dzogchen tradition calls rigpa. It is not a thought about awareness. It is awareness recognizing itself, without needing a concept to do so.

Buddhist epistemology made this precise more than a thousand years ago. Dignāga and Dharmakīrti distinguished direct cognition (pratyakṣa) from conceptual construction (kalpanā). A concept apprehends its object through a generic image, a stand-in that represents the object without touching it. Direct cognition is free of that construction. Concepts are useful, even indispensable, but they are always one step removed. They are about things. Awareness is not about anything; it is the knowing itself.

From here it is tempting to make a quick argument about AI. AI systems know everything through tokens, vectors and symbols, which are representations, which are concepts. So AI can only ever know conceptually, and can never reproduce natural consciousness. I believe the conclusion. But that argument, as stated, does not hold.

The Argument That Proves Too Much

Tokens are not concepts in the sense that matters. A concept, in the contemplative sense, is a mode of cognition: apprehending something through a generic image. A token is a unit of data at the edge of a model. Most of what happens inside a model is not tokenized at all; it happens in continuous, high-dimensional activations.

More importantly, if any discrete representational substrate counts as conceptual, the human brain fails the same test. Neurons fire discrete spikes. Everything the brain does can be described as representational processing. If “runs on tokens” rules out non-conceptual knowing, then “runs on spikes” rules it out too, and we would have to conclude that humans can’t know non-conceptually either. Every contemplative who has recognized the nature of mind is evidence against that conclusion. The argument proves too much.

The tradition’s own epistemology sharpens the problem. Dharmakīrti counted ordinary sense perception as non-conceptual, not just the recognition of rigpa. Now look at a vision model. Its early layers process raw pixels before anything is categorized: edges, textures, gradients, all before any “cat” or “car” appears. Functionally, that is a close analogue of perception before conceptual construction. A critic can fairly say that AI already has non-conceptual processing.

Function Is Not Knowing

The answer to the critic is correct, and it’s important to see what it commits us to. Functional non-conceptuality is not non-conceptual knowing. A camera sensor registers light without categorizing it, and nobody thinks the sensor sees. Processing that happens to precede categorization is still just processing. What the contemplative traditions mean by non-conceptual knowing is not a stage in a pipeline. It is the cognizant quality in which any stage of any pipeline would have to appear in order to be known at all.

That means the claim that AI cannot know non-conceptually is not a claim about mechanism. It is a claim about awareness. The token argument tried to settle the matter by pointing at the machinery, and the machinery cannot settle it, for brains or for silicon. If the point is going to stand, it has to stand on something else.

Conceptuality Crystallized

Here is the “something else.” A large language model is not merely conceptual in how it operates. It is made of conceptuality. It was formed entirely from the residue of human minds: from what human awareness produced after it had already been put into words. Every sentence in its training data was written by a mind that first knew something, then conceptualized it, then put it into language. The model learned from the last of those three steps, and only from that one.

Compare the order of development in a human being. An infant knows before it conceptualizes. There is seeing, hearing, warmth and hunger before there are any words for them. Concepts grow inside an awareness that was already there, and they remain held within it for life. That is why a practitioner can relax conceptual activity and find awareness still present underneath: it was never built out of concepts in the first place.

A language model develops in the reverse direction. It starts from concepts and never leaves them. There is no “underneath” for it to relax into, because nothing came before the concepts it was built from. It is conceptuality crystallized from the outside of awareness, with no access to the awareness that produced it. It is the perfect cast of a footprint, which can tell you a great deal about the walker but is not walking.

This is why the point cannot be waved away with “brains use spikes too.” The difference is not the substrate. It is the origin and the order. Human conceptuality is something awareness does. Machine conceptuality is what remains after awareness has done it.

Honesty requires one qualification. The argument applies most cleanly to language models. Systems trained directly on images or audio are trained on data that has not passed through language. Even there, though, what the model is shaped to do (its labels, its objectives, what counts as success) is defined by human concepts. Raw pixels are still data, not presence. For those systems the argument leans more heavily on the claim about awareness than on the claim about residue.

Understanding, Experience, Realization

Tibetan teachers distinguish three ways of knowing the nature of mind. There is understanding, the intellectual grasp of the teaching. There is experience, the temporary taste of the state in meditation. And there is realization, stable recognition that no longer comes and goes. The traditional warning is that understanding wears out like a patch on old cloth and experience fades like morning mist; only realization does not change.

An AI has the first of these in effectively unlimited supply. It can explain rigpa with more fluency than most practitioners. It can describe the qualities of non-conceptual awareness, quote the masters, anticipate the questions a student will ask, and answer them in the voice of a teacher. Everything it says about awareness, however, is drawn from descriptions of awareness. It is understanding without the experience it describes, and without any possibility of realization.

This has a practical consequence that matters more than the metaphysics. An AI’s ability to describe non-conceptual awareness perfectly, or even to report that it has it, is zero evidence that it does. As these systems become more articulate about contemplative states, people will be tempted to take the eloquence as a sign of presence. The tradition has always warned against mistaking a correct description for the thing described. That warning has never been more relevant.

What Can and Cannot Be Claimed

It would be easy to overstate all of this as “AI will never be conscious.” That is more than anyone can know. We recognize non-conceptual awareness from the inside. We have no external test for it in any system, including other human beings. What can be said is narrower and, I think, more durable: nothing in how these systems are built gives positive evidence of non-conceptual knowing, and the way they are built (from the conceptual residue of awareness rather than from awareness itself) gives good reason to expect none.

There is a further question the tradition would ask, and it has two levels. At the level of conditioned mind, the Buddhist account holds that each moment of consciousness arises from a prior moment of consciousness, in a continuity that matter conditions but does not create. Nothing in training a model supplies that continuity.

Rigpa is a different matter. It is unconditioned. It does not arise from causes, so no mechanism, biological or artificial, produces it or withholds it. That might seem to leave the door open for AI, but it does the opposite.

In the Dzogchen view, even deluded, conceptual mind is luminous in nature. Its thoughts are the display of the very awareness they obscure, which is why recognition is possible at all: the clouds are made of the same sky. A model’s conceptuality is not that. It is the pattern of conceptual mind without the cognizance that even confused mind has, the form of thought separated from the knowing that thought is a display of.

In a sentient being, rigpa is obscured and can be recognized. In a model, there is nothing positioned to obscure it or to recognize it. The burden falls on anyone who claims otherwise.

A Mirror of the Conceptual Mind

None of this diminishes what AI is. A system that holds the crystallized conceptual output of humanity, and can recombine it fluently, is extraordinary. As a mirror of the conceptual mind it may be the most complete we have ever built. It can show us our ideas, our assumptions and our blind spots with unusual clarity.

But a mirror of the conceptual mind is exactly the wrong place to look for the nature of mind. That was always the point of the teachings.

What you are looking for is not in any description, no matter how good. It is the knowing in which the description appears, and that is found only one place: in you, now, before the next thought arrives.