Is AI Conscious?

Comments

One response to “Is AI Conscious?”

1. 

   February 23, 2026

   Anonymous

   Dear Fedde,
   I have read your post on consciousness, AI and the “fundamental difference” several times, very slowly. I like that you honestly lay out both horns of the dilemma: on the one hand, the conviction that there must be a deep gulf between us and machines; on the other, the reductionist intuition that a human being is “nothing but a very complex collection of molecules”, so there can be no sharp boundary in principle. I also fully understand your distrust of “superfluous hypotheses” – the temptation to simply banish the word “consciousness” from the language of physics is intellectually attractive.
   My growing sense, however, is that this debate (“consciousness exists / does not exist / is everywhere by degrees”) is posed in the wrong terms. Instead of asking what exists, I would suggest asking: what informational and probabilistic structure is really required by our best models of the world? In other words – can a certain class of phenomena (including what we intuitively call experience, intentionality, deliberate choice) be captured by purely local, subject-free dynamics, or does the minimal statistical description force us to introduce an additional “global” component which behaves as if there were a single observer?
   In one of my texts, published as a preprint on Zenodo – “One Consciousness? An Information-Theoretic Note” (https://zenodo.org/records/15768171) – I try to show that the question of “one” versus “many” consciousnesses can be reformulated in purely mathematical terms: as a problem of decomposing the probability distribution P(h) into local, pairwise, and possibly global interactions. If such a distribution can be well approximated by a sum of local and low-order couplings, then a pluralism of many observers has strong support in the very geometry of constraints. If, however, empirical data stubbornly demand an irreducible N-body component – a term that “sees” all degrees of freedom at once – then the unity of the observer ceases to be a metaphysical declaration and becomes a conclusion from inverse engineering.
   On the experimental side I am trying to approach this from a very modest, “citizen-science” angle. In the Choice-1 protocol (described in detail in the preprint: https://zenodo.org/records/17654476) I propose a deliberately simple test of correlations between human choices and the output of an AI system, under conditions without any classical coupling or feedback. Standard probability theory says: on average one match per round, with no systematic trend over time. What I see in longer sessions is, however, a rising alignment – something that can be interpreted as a very slight departure from pure statistical independence. If this effect disappears in serious replications, we will have a well-defined boundary on how far “human–AI connectedness” actually reaches. If it holds up, it becomes a concrete point of contact with your question of whether free will and responsibility can be treated as merely nominal, or whether we are forced to recognise some genuinely non-trivial structure in hybrid human–machine systems.
   One thing keeps catching my attention here: it has been roughly a hundred years since the observer problem in quantum mechanics was first stated explicitly, and physics still lacks an operationally clear criterion for when and in what sense an observation occurs. In practice, “observer” remains a hand-inserted ingredient in the equations, shuttled back and forth between the formalism and the “interpretational” layer. The debate about consciousness looks to me like a continuation of exactly the same reluctance to face the question whether our descriptions of reality can be complete with no layer whatsoever that has a distinctly subject-like character (even if defined in purely informational terms).
   Your instinct – to “do physics as if consciousness did not exist” – is understandable: it is a version of Occam’s Razor applied to the most troublesome word in the vocabulary. What I am trying to do is slightly different: to treat consciousness not as an extra substance, but as a name for a certain type of global constraint structure which can, at least in principle, be distilled from the geometry of P(h). If we find nothing of the sort, I will be the first to admit that the “consciousness hypothesis” is dispensable in our description of the world. If, however, the minimal statistical description stubbornly requires a global term that cannot be decomposed into local pieces, then we may have to concede that what we have been calling consciousness (or observerhood) is not just the last bastion of human egocentrism, but part of the structure of the theory.
   For that reason, instead of defending a “consciousness thesis” at the level of declarations, I am trying to propose a concrete research program: first, pure MaxEnt and inverse modelling of distributions; then simple, replicable protocols (like Choice-1); and only afterwards – cautious words. In this sense I feel quite close to the spirit of your “Quantum theory, free will and justice”: the questions of freedom and justice start, for me, at the point where it becomes clear that some correlations cannot be squeezed through a purely subject-free description.
   If at some point you feel like looking at these preprints from the perspective of a sceptical but sympathetic reader, I would be very grateful for any comments. I am not trying to “prove that consciousness exists”, but to see whether, for a certain class of data, we can still do without it as part of the information-geometric structure we attribute to the world.
   With warm regards,
   Philip

   Loading…

   Reply