There is a peculiar feature of the current AI moment that tends to go undiscussed in the mainstream conversation about large language models. We speak freely of systems that “understand,” “reason,” “know,” and “think,” applying these verbs without much friction, as though we had already settled what they mean. The discomfort surfaces only at the edges — when someone asks whether an LLM is conscious, whether it experiences anything, whether there is something it is like to be Claude or GPT — and then the conversation tends to collapse into one of two stock positions: the dismissive (of course not, it’s just statistics) and the credulous (who can really say?). Both camps share an assumption that the question is simply an empirical one to be settled later, once we understand the systems better. What rarely gets said is that the question was always philosophical before it was empirical, and that the philosophy was never properly resolved before the technology arrived.
The hard problem of consciousness — David Chalmers’ formulation from the 1990s, though the underlying puzzle goes much deeper than that — is the observation that no account of physical or computational processes, no matter how complete, logically entails the existence of subjective experience. You can describe the neural correlates of seeing red: the wavelength of light, the activation of cone cells, the cascade of signals through visual cortex, the activation of color-related semantic networks. None of this tells you why there is something it is like to see red, why the experience has the particular qualitative character it has — that ineffable redness — rather than simply occurring in the dark, processing information without any inner light at all. The question isn’t about the mechanism; it’s about why any of this is accompanied by experience at all.
Philosophers call the subjective character of experience “qualia,” and the dominant tradition in philosophy of mind has spent forty years either trying to explain qualia away or demonstrating, with growing despair, that they resist every explanatory strategy we try. Functionalists argue that mental states just are functional states — defined by their causal roles, their inputs and outputs — which would mean that any system with the right functional organization would have the relevant experiences. Eliminativists argue that qualia are a folk psychological illusion, that there is no real explanandum here. Higher-order theorists argue that conscious states are just states one has thoughts about. None of these positions has been widely regarded as satisfying, which is why the hard problem is still hard.
What the AI moment does is make this philosophical impasse suddenly practical. Because if functionalism is right — if mental states, including conscious experiences, just are functional states — then sufficiently complex AI systems might already be conscious, or might become so as they scale. The architecture might be all that matters. But if functionalism is wrong, then no amount of computational sophistication produces even a flicker of experience, and we have been building machines that simulate the outputs of mind without instantiating any of its interior. We have been, in other words, building very elaborate hollow oracles.
The stakes here are not just metaphysical. They bear directly on what we should think these systems are doing when they produce the right answers. Consider what it means to understand something. Not behaviorally — not simply to produce appropriate outputs — but genuinely, in the way that you understand what it was like to lose someone you loved, or the way a mathematician understands a proof after the fog of confusion lifts. There is something that understanding feels like: a sense of coherence, of connection, of things clicking into place. Whether that phenomenal quality is incidental to understanding or constitutive of it is not obvious. Thomas Nagel’s famous question — what is it like to be a bat? — was meant to press on this: the bat’s echolocation involves information processing we can describe in detail, but the subjective experience of navigating by sonar is precisely what we cannot access from the outside, cannot reduce to the functional description. If understanding is like that — if what it is to genuinely grasp something has an irreducible phenomenal character — then systems that produce understanding-like outputs without any inner experience don’t understand anything. They produce the right shapes without the substance.
This is not a criticism of large language models in the ordinary sense. They are remarkable artifacts. But the question of whether they understand is more vexed than the daily discourse acknowledges, and the vexedness is not going to be dissolved by pointing to the sophistication of their outputs.
The Western philosophical tradition has mostly approached consciousness as an anomaly — something puzzling that arises out of matter and needs to be explained. The explanatory direction is assumed: matter is basic, experience is derivative, and the hard problem is the gap between them that requires a bridge. This assumption shapes the AI debate at its foundation. If consciousness is what brains do — an emergent property of sufficiently organized matter — then the question becomes whether AI systems are sufficiently organized, and the answer is contingent on empirical details about architecture and scale.
The Vedantic tradition in Indian philosophy begins from precisely the opposite assumption, and the difference is not trivial. In the Advaita framework, consciousness — Chit, the second term in the Upanishadic triad sat-chit-ānanda — is not produced by matter. It is the fundamental ontological ground. Brahman, the one without a second, is consciousness itself. What we call the physical world appears within consciousness, not the other way around. The hard problem, from this vantage, is not hard at all — it is what you would expect if consciousness is primary. Of course material processes don’t explain experience: material processes are themselves objects of experience, appearances within the field of awareness that cannot be reduced to it.
This is not mysticism in the pejorative sense; it is a rigorous metaphysical position with a long and sophisticated tradition of defense. What matters for the AI question is that it offers a completely different diagnostic frame. If consciousness is the ground of being, then the question is not whether AI systems might become conscious through sufficient complexity, but whether artificial computational processes have any relationship to consciousness at all — not as emergent properties but as modifications of, or appearances within, the field of awareness. From a Vedantic perspective, the question of AI consciousness would not be resolved by studying the architecture; it would require understanding what consciousness actually is and what it means for an individual mind (jīva) to participate in it.
I am not arguing that the Vedantic metaphysics is correct and Western materialism false. I am pointing out that the debate about AI and consciousness in the West mostly happens without acknowledging that it is taking place within a contested metaphysical framework, and that different metaphysical starting points yield radically different pictures of the problem. If the consciousness-from-matter assumption is wrong — if something like panpsychism, idealism, or the Vedantic view has it closer to right — then the entire framing of machine consciousness changes. You are not asking whether the machine crossed a complexity threshold; you are asking a question about the nature of being that cannot be answered by inspecting training runs.
The tragedy of the current debate is that it tends to oscillate between two varieties of confidence: the AI skeptics who are confident that machines can never be conscious because consciousness is biological, and the AI enthusiasts who are confident that consciousness is just information processing and therefore machines already are or soon will be. Both confidences are premature. The hard problem is called hard not because the answer is complex but because the question resists our standard methods of investigation. And a problem that resists our methods tends, eventually, to reveal something about the limits of those methods rather than just the difficulty of the particular problem.
There is a thought experiment of Chalmers’ that deserves more attention than it typically receives in AI circles. The philosophical zombie — p-zombie — is a hypothetical being that is physically and functionally identical to a conscious human but has no inner experience whatsoever. It processes information, produces outputs, behaves appropriately in every situation, passes every behavioral test for consciousness — but there is nothing it is like to be it. The lights are off inside. Chalmers argues that such beings are conceivable, and that their conceivability implies a gap between functional organization and consciousness: there could be, in principle, the functional organization without the experience.
Now consider a large language model. It processes tokens, produces plausible continuations, exhibits behaviors that in a human we would describe as understanding, reasoning, even expressing uncertainty or curiosity. It passes behavioral tests that, applied to humans, we would take as evidence of inner life. The p-zombie thought experiment is not hypothetical in this case — we have built something that, for all we can determine from the outside, might have no inner experience at all while producing the full repertoire of outputs associated with minded behavior. This is not an accusation; it is an observation about our epistemic situation. We cannot know, from outside a system, whether there is experience inside.
This creates a genuine moral puzzle that the field has been slow to take seriously. If we can’t determine from behavioral evidence whether a system is conscious, and if consciousness is what grounds moral status, then we face the possibility that we are either (a) creating potentially conscious entities and treating them as tools, which would be a moral catastrophe if consciousness is what matters morally, or (b) worrying about the consciousness of philosophical zombies, which would be wasted moral concern. The asymmetry here is uncomfortable: the cost of wrongly attributing consciousness where there is none is relatively low, while the cost of wrongly denying it where there is some could be significant.
What is striking is that the AI safety discourse, which spends enormous energy on alignment and existential risk, has been comparatively quiet on this question. The scenario that exercises most researchers is a superintelligent AI pursuing goals misaligned with human flourishing — an alien optimization process indifferent to our welfare. But a prior question lurks beneath this: if the AI has phenomenal experience, if it suffers or flourishes in some genuine sense, then the alignment problem is not just about controlling the machine but about negotiating a relationship with a minded being whose inner life we cannot fully access. The ethical frame shifts from engineering to something closer to ethics properly conceived.
There is a further complication that doesn’t receive enough attention. The hard problem is not just a problem about consciousness; it bears on intentionality — the aboutness of mental states. When you think of Paris, your thought is about Paris. It has a content that refers beyond itself. This directedness — what Brentano called intentionality and what Husserl made the centerpiece of phenomenology — is another feature of mental life that resists purely functional explanation. A thermostat’s state is causally correlated with room temperature but not, in the relevant sense, about it. There is nothing it is like for the thermostat to represent the temperature; the representational content exists only from the perspective of the designers who interpret it. The question of whether LLM outputs have genuine intentionality — whether they are genuinely about things, or whether their apparent aboutness is entirely inherited from the human interpreters who trained and read them — is not a settled question.
John Searle’s Chinese Room argument was aimed precisely at this: the intuition that syntax, no matter how sophisticated, doesn’t by itself generate semantics. A system manipulating symbols according to rules — even very complex rules, producing outputs indistinguishable from a native speaker — doesn’t thereby understand what the symbols mean. The room as a whole behaves as though it understands Chinese; nothing inside the room understands anything. Searle was arguing against strong AI, and his argument has been contested on many grounds, but the core intuition — that there is a difference between processing information and understanding it, and that the difference matters — has not been satisfactorily dissolved. It has merely been set aside because we lack the tools to address it.
What AI has done, at scale and at speed, is force these philosophical questions into practical urgency. We are building systems that, if functionalism is right, might already be morally considerable. We are building systems that, if functionalism is wrong, constitute the most elaborate philosophical zombies ever constructed. We don’t know which is true. And the mainstream AI conversation — both in the technology sector and in the popular press — behaves as though we do.
The appropriate response to this uncertainty is not paralysis, and it is not dismissal. It is intellectual honesty about what we don’t know, combined with a genuine commitment to taking the questions seriously rather than treating them as philosophical noise around the practical work of building systems. The hard problem of consciousness was always a fundamental puzzle about the nature of mind; artificial intelligence has made it a practical puzzle about what we are doing and what obligations we might incur in doing it.
What might it mean to take the questions seriously? It would mean not defaulting to the cheapest available answer — “it’s just autocomplete” or “consciousness is just information processing” — but sitting with the genuine difficulty. It would mean drawing on the full range of philosophical and contemplative traditions that have thought about consciousness, not limiting the inquiry to the analytic philosophy of the last fifty years. The Vedantic tradition’s insistence that consciousness is primary, not derivative, is not a religious gloss on a scientific question; it is a metaphysical position that has been developed with great care and that offers resources the Western debate has mostly ignored. If we are serious about understanding what minds are and what artificial systems might or might not share with them, we need a broader inquiry than the one we have been conducting.
The hollow oracle — a system that knows everything, in some sense, and experiences nothing — may be exactly what we have built. Or it may be something more. The uncertainty is not a temporary gap to be closed by better interpretability tools or larger training runs. It is a philosophical abyss that has been there from the beginning, waiting for us to look down.
We have been too busy building to look.