Scott Alexander has a nice review of a new paper called “Identifying indicators of consciousness in AI systems”. (This paper, which appears in Trends in Cognitive Science, is by Patrick Butlin, me, and many co-authors across neuroscience, AI, and philosophy). In his post, Scott divides theories of consciousness into three buckets:

Physical: whether a system is conscious depends on its substance or structure.

Supernatural: whether a system is conscious depends on something outside the realm of science, perhaps coming directly from God.

Computational: whether a system is conscious depends on how it does cognitive work.

This post is about how this division is both useful, and not quite right (one example: the first category shouldn’t be called “physical”). Along the way, the limitations of Scott’s division will point the way to a better way of carving up different views of consciousness.

Let’s start with what’s good about Scott’s three-way division into physical, supernatural, and computational: it leads to a clear and correct explanation of what we do in the paper, i.e. look at computational theories of consciousness and using them to derive indicators of consciousness that we can look for in AI systems.

What does this methodology pre-suppose? On what views does it make sense? Scott’s categories help us see that we basically need consciousness to be computational and lawlike:

(a) consciousness is not inherently tied to biological substance [in contrast with theories that Scott calls “physical”]

(b) there are some lawlike principles governing consciousness [in contrast with what Scott calls “supernatural”].

If either of those conditions fails, then AI consciousness is either impossible or unknowable. As Scott puts it: “If consciousness depends on something about cells (what might this be?), then AI doesn’t have it. If consciousness comes from God, then God only knows whether AIs have it.”

Scott correctly notes that we don’t take ourselves to have proven that consciousness is computational and lawlike. Rather, we think that consciousness has a decent chance of being computational and lawlike, and that we can make progress within that region of credence space. We are making a conditional claim: IF consciousness is computational and lawlike, then X and Y are computational indicators of consciousness. We aren’t begging the question against (e.g.) biological views of consciousness; we’re not even trying to argue against those views.

With all that said, there are some confusing aspects of Scott’s division: by conflating “physical” and “biological” in the first category, and relatedly making “computational” a contrast with “physical”, Scott obscures some important axes that we should actually keep separate. Here’s how I think we should divide things:

Axis 1: Metaphysics. Is everything, including consciousness, fundamentally material? Or is there something non-material? Materialism says yes, everything is fundamentally material. Dualists say no—reality fundamentally includes matter but also some non-material properties (property dualism) or non-material stuff (substance dualism). Panpsychists also say no, but in a more exotic way—the fundamental stuff is somehow conscious-y.

Axis 2: Lawfulness. Are there lawlike regularities between states of matter and states of consciousness, such that a science of consciousness is possible? Importantly, all of the positions on axis 1 can answer “yes” to this. Dualists can think that when God wrote the laws of physics (as it were), he also wrote some additional laws of nature that link matter with consciousness. Materialists will think that God rested right after he wrote the laws of physics, having completed his task. But materialists and dualists can both in principle agree that we can look for the laws that govern what physical stuff correlates with consciousness.

Axis 3: Level of description. At what level are these correlates of consciousness specified—biological implementation, or something more abstract like computation/function? I think this is where the real action is for AI consciousness, and where most of the disagreement lies. A “biological naturalist” thinks that the laws linking matter to consciousness must be specified at the level of neurons, carbon-based chemistry, or some other distinctly biological property. A computational functionalist thinks the laws can be specified at a more abstract level—at the level of algorithms and information processing—such that consciousness can be realized in computer chips as well as living neurons.

The key point is that these three axes are independent. Scott conflates metaphysics and level of description by contrasting “physical” and “computational”. But computations can certainly be construed as physical! They’re realized in physical substrates, and one way of being a physicalist about consciousness is to explain it in terms of computations. Computational functionalism pairs naturally with physicalism (though dualists can be computationalists too); it just denies that consciousness requires any particular physical substrate. So the real contrast isn’t physical vs. computational, it’s biological vs. computational.

And this independence is why David Chalmers—the philosopher most associated with the “hard problem” and property dualism—is a co-author on this paper. People sometimes find this puzzling. Shouldn’t a property dualist be skeptical of projects like this? If consciousness is something over and above the physical facts, why would examining AI’s computational structure tell us anything?

But that would be to conflate axis 1 with axis 3. Chalmers talks about this in his paper “The Singularity: A Philosophical Analysis”:

I have occasionally encountered puzzlement that someone with my own property dualist views (or even that someone who thinks that there is a significant hard problem of consciousness) should be sympathetic to machine consciousness. But the question of whether the physical correlates of consciousness are biological or functional is largely orthogonal to the question of whether consciousness is identical to or distinct from its physical correlates. It is hard to see why the view that consciousness is restricted to creatures with our biology should be more in the spirit of property dualism!

In other words: your position on axis 1 (the metaphysical question) doesn’t determine your position on axis 3 (the level of description / substrate question). A property dualist can perfectly well think that consciousness lawfully accompanies certain computational structures regardless of whether they’re implemented in neurons or silicon.

So the paper’s methodology is informative for anyone who thinks there are lawlike regularities governing consciousness (axis 2: yes) and thinks those regularities are specified at the computational level rather than the biological level (axis 3: computational)—regardless of whether they think consciousness is identical to or distinct from its physical correlates (axis 1: either way).

That’s the region of credence space where we can and should be looking at AI architectures for computational indicators.