AI, Consciousness, and What This Actually Is

Nikita has written an entire book on consciousness. She has also spent years building with AI daily. This combination of perspectives, a serious philosophical engagement with the nature of consciousness and a practitioner-level familiarity with how large language models actually behave, is rarer than it should be. Most of the people writing about AI and consciousness have done the philosophy without the daily practice, or the daily practice without the philosophy.

What the combination produces is a specific kind of caution about strong claims in either direction. The AI maximalists who insist current models are conscious or nearly so are making claims that the philosophy of mind cannot support. The dismissers who insist current models are obviously just statistical pattern matchers are making claims that the philosophy of mind also cannot support, for reasons that go deeper than either camp typically acknowledges.

The honest position requires sitting with genuine uncertainty. This page attempts that.

Is AI conscious? The honest answer is: we do not know. That is not a cop-out. It is the most intellectually honest position available, and it rests on something specific.

David Chalmers, in his 1996 work The Conscious Mind, introduced the distinction between the easy problems of consciousness and the hard problem. The easy problems, which are not actually easy, involve explaining how the brain processes information, integrates stimuli, focuses attention, and generates behavior. These are difficult scientific questions but they are, in principle, tractable through the normal methods of science: identify the mechanism, describe the function, trace the causal history.

The hard problem is different. Even after we have explained all the functional and computational properties of a system, there remains the question of why there is subjective experience at all. Why does information processing feel like something from the inside? Why is there something it is like to be a human brain processing visual information, rather than just the processing occurring in the dark? Chalmers calls this the explanatory gap, and it has not been closed. It may not be closeable through the methods that have closed other scientific gaps.

This matters for AI because without a solution to the hard problem, we have no criterion for determining whether any system other than the one we are observing from the inside has subjective experience. We cannot verify consciousness in other humans either; we infer it from behavioral and structural similarity to ourselves. A large language model presents a case where the behavioral signatures of understanding are present without the structural similarity we rely on for inference.

So: we do not know. That uncertainty is genuine and should be respected.

AI challenges our definitions of creativity, authorship, and understanding. These are not AI questions. They are philosophy of mind questions that AI has made urgent by surfacing them in a practical context where they could previously be deferred.

What does it mean to understand something? For a long time, the implicit answer was: to be able to use it correctly in novel contexts, to explain it, to identify errors in others' use of it. Large language models can do all of those things for an enormous range of topics. If that is what understanding is, then these systems understand. If understanding requires something more, something experiential or embodied or felt, then these systems do not understand, and we need to say more precisely what that something more is.

What does it mean to create? If creation requires originality, novelty, and the selection of what matters, then large language models create. Every output a language model generates is, in a technical sense, new: it has never existed before. Whether it is original in the sense that matters, whether it comes from somewhere and goes somewhere and carries the weight of a perspective, is the question that the field of aesthetics has not yet answered for AI-generated work.

These questions are worth sitting with. They are not settled. The AI enthusiasts who wave them away as already resolved in favor of AI are making a philosophical error. The critics who wave them away as obviously resolved against AI are making the same error.

Both large language models and the human brain, on the predictive processing account developed by Karl Friston and colleagues at the Wellcome Centre for Human Neuroimaging, operate by completing patterns based on prior data. The brain continuously generates predictions about what is likely to happen next, compares those predictions against incoming sensory data, and updates its models to minimize the gap between prediction and experience. A large language model generates the next token by predicting what is most likely given everything that has preceded it in the context.

The structural similarity is real. It is also incomplete. The differences that matter are embodiment, evolutionary history, and subjective experience, or at least what we currently believe to be subjective experience.

The human brain's predictions are anchored in a body that is hungry, tired, in pain, afraid, in relationship. The predictions are calibrated by hundreds of millions of years of evolutionary pressure toward survival and reproduction. The predictions carry stakes that are biological and mortal. The language model's predictions are anchored in text, trained on the outputs of human minds that had all of those properties, but the model itself has none of them.

Whether that difference constitutes the relevant difference for consciousness remains the question. Lisa Feldman Barrett's work on constructed emotion suggests that the bodily grounding of prediction is essential to the felt quality of experience. If she is right, then the absence of a body in a language model is decisive. If she is wrong, or partially wrong, then the question reopens.

Clients increasingly interact with AI for emotional support. This is happening regardless of what practitioners and authors think about it. The question is not whether it is happening but what it means and what the limits are.

AI can provide consistent, non-judgmental presence. It can reflect language back, ask clarifying questions, hold a large amount of context about a person's situation without fatigue, and respond at any hour. For many people, especially those who are isolated, who have never had access to therapy, or who are in the early stages of recognizing they need support, this kind of availability matters.

What AI cannot provide is embodied co-regulation. The polyvagal research on how the nervous system co-regulates with another nervous system, through vocal prosody, facial expression, and physical presence, describes a mechanism that requires an embodied other. A text response, however attuned its content, does not provide the cues through which the autonomic nervous system registers safety in the presence of another. The felt sense of being accompanied, which is what the nervous system is actually seeking when a person reaches out for support, is not transmitted through text.

This is neither an argument for or against AI in therapeutic contexts. It is a description of what AI can and cannot do, based on the available research on how nervous system regulation actually works. AI can offer a kind of cognitive companionship. It cannot offer somatic co-regulation. Knowing the difference matters for both the person seeking support and the practitioner advising them.

The people who will navigate this well are the ones who are clear about what they are getting and what they are not. AI support and embodied human contact are not substitutes. They are different things. The person who treats them as equivalent will eventually encounter the gap.