Seth’s case that artificial consciousness is unlikely, on current development trajectories, comprises both a negative half (critiquing computational functionalism, which suggests artificial consciousness comes for free with artificial intelligence), and a positive half (defending biological naturalism, which makes artificial consciousness depend on artificial life).

While I have some sympathy with both halves, I won’t try to evaluate either computational functionalism or biological naturalism here. Rather, I will raise two objections, both suggesting that, even if we reject computational functionalism, artificial consciousness is still likely, albeit in strange and unfamiliar forms. This points towards a question lurking in the background: how sparse or abundant should we expect consciousness to be?

My first objection concerns substrate-independence (or substrate-flexibility), the idea that the very same conscious process could be implemented in basically any sort of material, if it is functionally isomorphic to a brain. Seth points out that if this is true, then making a silicon computer functionally isomorphic to a human brain, in some coarse-grained sense, should suffice for artificial consciousness. But he often writes as though the flipside of this is that if substrate-independence fails, then systems with a radically different substrate than us would completely lack consciousness. But of course, strictly all that follows from rejecting substrate independence is that a system with a radically different substrate would have to differ from us in respect of consciousness. That’s compatible with it still being conscious, and compatible with its consciousness resembling ours in lots of ways. It just means that it cannot be exactly the same: A difference in substrate might systematically alter the system’s phenomenal qualities, phenomenal structure, agency, hedonic valence, or something else, without depriving it of consciousness altogether.

My second objection then concerns how we should decide between these possibilities. Suppose we find ourselves dealing with a system that is both very different from us in substrate and very similar in function. In particular, suppose that it makes claims and demands, talks about what it wants and likes and dislikes, even describes its internal states as “pleasant” or “unpleasant,” and that we are in this situation without much confidence in computational functionalism, or biological naturalism, or indeed any particular theoretical framework. We’re profoundly unsure what makes something conscious, but here’s a being before us that acts like it’s conscious and claims to be conscious: do we default towards accepting or rejecting its self-ascriptions?

It seems to me that our default here should be acceptance: we need a compelling argument to doubt a system’s self-ascriptions of consciousness, but not to accept them. This default might be justified by precautionary pragmatic reasons (for discussion, see e.g. Long et al., Reference Long, Sebo, Butlin, Finlinson, Fish, Harding, Pfau, Sims, Birch and Chalmers2024), or by the “Copernican” argument that we should not assume our own substrate is somehow uniquely privileged (see Reference Schwitzgebel and PoberSchwitzgebel and Pober Ms.), or by reflection on the fact that we obviously have ascribed consciousness to other humans (and animals, and historically many other things) without the foggiest idea of the correct theory of consciousness.

Of course, that default might be cancelled in various ways. We might look at the system’s history and find that it was deliberately designed to mimic human discussions of consciousness, or to game out human tests for consciousness, and reason that this cancels out the evidential value its words and actions would otherwise have (Andrews & Birch, Reference Andrews and Birch2023, Reference Schwitzgebel and PoberSchwitzgebel and Pober Ms.). For my part, I think this sort of history-based defeater is currently the most compelling reason to doubt the consciousness of present or near-future AI. But note that this applies to both biological and silicon-based systems, if they have the same kind of history. So it does not directly intersect with the dispute over functionalism and naturalism.

We could also override the default by appealing to a theory of consciousness. If we were strongly convinced of Seth’s biological naturalism, we might reject an AI’s self-ascriptions of consciousness on the basis that it is not biologically alive. But justifying such denial surely requires high confidence in biological naturalism, not just thinking it the best current theory. The relevant comparison would not be with specific rival theories like computational functionalism, but with a general stance of humble agnosticism. Because if we’re agnostic about the right theory of consciousness, then it’s hard to justify theory-based rejections of a system’s claims to be conscious.

I think all of this taps into an underlying question: how abundant or sparse should we expect consciousness to be? Do we live in a world full of consciousness, where systems with different functional organisations are conscious in different ways, and systems built out of different substrates are conscious in different ways? Or is “consciousness” the name for something odd and rare, something unique to humans and beings that match humans sufficiently closely in some key respects (in organisation, in substrate, or in something else)?

In a world of abundant consciousness, even if we reject substrate-independence, we should still default to taking differences of substrate as indicating differences in consciousness, not its absence. (To do otherwise would be human exceptionalism.) Whereas if we assume a world of sparse consciousness, then it makes more sense to deny consciousness to systems that, even if they resemble us in some ways, differ from us in any fundamental way. (To do otherwise would be anthropomorphism.)

For my own part, I think the best-supported current theory is panpsychism, an extreme form of the abundant view (for arguments, see Chalmers, Reference Chalmers, Nagasawa and Alter2015, Seager, Reference Seager2006). Panpsychism reframes the question of AI consciousness: instead of asking “is this system conscious or not?,” it suggests we ask “what sort of consciousness is present in this system?,” and thus clarify what forms or dimensions of consciousness we’re interested in (cf. Birch, Schnell, & Clayton, Reference Birch, Schnell and Clayton2020). Do we want to know whether AI has valenced experiences? A unified perspective? Something else? This outlook pushes us away from deciding between functionalism and naturalism, as theories of whether a system is conscious, and towards asking how different factors, organisational or substrate, determine the dimensions and parameters of consciousness.