Anil Seth’s 2025 work on conscious artificial intelligence is one of the more honest positions in the field, which is exactly why it deserves a careful answer rather than a dismissal. I have written one. The full commentary is on Zenodo, and I will get to the link; first the disagreement, because a link without an argument is just advertising.

Seth argues, in broad strokes, that consciousness is a biological phenomenon — that it is bound up with being a living system, and that computation or function alone is unlikely to be sufficient for machine consciousness. The label he borrows from Searle is biological naturalism. The intuition underneath it is a good one: consciousness probably depends on real causal powers, on the specific physical properties of the system that has it, and not merely on some abstract input-output profile that any sufficiently clever simulation could reproduce. On that last point Seth is right, and most functionalists are too quick to wave it away.

Where I part ways with him is on what follows from that intuition. “It needs real causal powers” is true. “Therefore it needs to be biological” does not follow. Those are two different claims, and the gap between them is where the actual theory has to go.

The functionalism objection, and why criticality answers it

The standard worry about functional accounts of consciousness is that they are too permissive. If consciousness is just the right function, then in principle a lookup table, a sufficiently large pile of beer cans, or a bureaucracy passing notes could be conscious, which most of us take to be a reductio. Seth’s response is to anchor consciousness back in biology, where the causal powers are obviously real. That solves the permissiveness problem by fiat — but it also smuggles in an empirical bet (carbon matters, silicon cannot) that nobody has actually cashed out.

The Four-Model Theory takes the other route. It accepts the premise — consciousness needs genuine causal powers, not just an abstract functional profile — and then specifies what those powers are. Consciousness, on this account, is a real-time, self-referential simulation maintained at dynamical criticality: a system modelling itself, at the edge between order and disorder where its own dynamics become maximally sensitive and integrated.

That is the move worth noticing. Criticality is not “any computation will do.” It is a concrete, physical-dynamical requirement — a specific operating regime that a system either is or is not in, and which you can measure. So the Four-Model Theory is functional in the sense Seth distrusts, but it is not the naive functionalism his objection targets. It threads between biological essentialism on one side and lookup-table permissiveness on the other: it agrees that real causal powers are required, and it names them, rather than outsourcing them to “being alive.”

The consequence is the one Seth’s framework is built to resist. If the requirement is self-referential simulation at criticality, then the substrate is — in principle — open. Not any substrate; a substrate that can hold itself at criticality and model itself. But that is a property of organization and dynamics, not of carbon. This is the same ground I covered in Can AI Be Conscious?, and the commentary sharpens it against Seth’s specific position.

A place where the two theories disagree on data

None of this would matter if it were unfalsifiable hand-waving, which is the usual fate of consciousness theories. So here is the empirical edge. The Four-Model Theory makes a prediction that diverges from standard predictive-processing accounts concerning psychedelic effects on anosognosic awareness — awareness of one’s own deficits. The two frameworks expect different things to happen there. That is a place where the disagreement is not philosophical but experimental, and experiments settle it. I keep the prediction at the level the theory licenses in the commentary; the point is simply that there is one, and that it cuts.

This is the difference between a position and a theory. Biological naturalism tells you where consciousness has been found. It does not tell you what would have to be true of a new system for it to be conscious — and so it cannot generate the requirement that would let you build, or rule out, machine consciousness. The Four-Model Theory states those requirements concretely enough to be wrong.

The commentary

Full text, with the argument laid out properly:

Further reading: the full Four-Model Theory paper is at https://doi.org/10.5281/zenodo.18669891, and the book-length treatment, The Simulation You Call I, is here.