Can We Test an AI for Consciousness? Part Three: The Meat Condition

Abstract

Parts One and Two covered a method for assessing AI systems for consciousness, and what it found. Both rest on one assumption: that performing the right computations is what makes a system conscious, whatever it happens to be made of.

Ned Block, writing in the same journal, argues that this assumption is doing enormous work and has not been earned. He proposes that consciousness may additionally require something below the level of computation — a biological condition he calls the meat condition.

His argument is not that machines cannot think. It is that our criteria for attributing consciousness contain a hidden fork, and that which way you take it decides the answer before any evidence arrives.

1.The Target

Block names the work in Part Two directly. He describes an influential paper on AI consciousness with nineteen authors, including a recipient of the 2018 Turing Award, and quotes its first stated assumption: that implementing computations of a certain kind is necessary and sufficient for consciousness.

This is computational functionalism, and Block's objection is precise. He does not claim it is false. He claims that current discussions do not take sufficiently seriously the possibility that implementing certain computations is not sufficient — because a subcomputational biological condition might also be required.

Subcomputational is the operative word. Block is not pointing at anything a computation does. He is pointing at the physical machinery that realises the computation, and asking whether that machinery has to be a particular kind of thing.

Theories of consciousness as currently formulated, he observes, are meat-neutral: they specify roles and relationships without saying what must fill them. That neutrality is usually treated as a virtue. Block treats it as an unexamined commitment.

2.The Fork

The most useful part of the argument is not the thesis but the structure Block exposes underneath it.

Nobody has direct access to consciousness in anything other than themselves. Every attribution to any other entity — another person, a dog, an octopus, a bee, a language model — is an extrapolation from the human case. The question is what you extrapolate on.

If you extrapolate on our computational properties, AI systems score well: they perform the relevant operations, often better than simple animals do. If you extrapolate on the subcomputational biological machinery that realises those properties in us, simple animals score well, because they are made of the same materials, and AI scores at zero.

Block's conclusion follows immediately and is genuinely uncomfortable: AI and simple animals are competitors when it comes to attribution of consciousness. Criteria that make a language model a candidate make an insect a poor one, and criteria that admit the insect exclude the machine.

This is the point at which the debate stops being about AI. Anyone who thinks bees might be conscious — a live question we cover elsewhere in this department — and also thinks language models might be, is holding two positions that pull in opposite directions and owes an account of how.

3.The Empirical Part

What lifts this above a philosophical stalemate is that Block proposes something to look at.

In 2023, a study in Nature supported comb jellies — ctenophores — as the sister group to all other animals, using gene linkage evidence. Ctenophores are relevant here for an unusual reason: at least at one stage of their life cycle they have an entirely electrical nervous system, with chemical synapses only where the system meets the environment, at sensory transducers and motor effectors.

Every other animal with a nervous system, including us, is electrochemical: signals propagate electrically within neurons, but communication between neurons is mostly chemical, through neurotransmitter release.

Block sets out three evolutionary scenarios for how this could have arisen, and notes that on all three, electrochemical processing appears to hold a functional advantage over purely electrical processing. And the animals we take seriously as candidates for consciousness are, without exception, the electrochemical ones.

From which he draws a research programme rather than a conclusion: we should be investigating what is special about electrochemical computation.

That is a testable proposal. It is also, notably, the same move Butlin and colleagues make — asking what feature of the substrate matters — applied one level further down than they apply it.

4.What This Does Not Establish

The meat condition is, in Block's own presentation, unspecified. He argues that such a condition might exist and that the field has not ruled it out. He does not identify it, measure it, or show that it obtains.

The evolutionary argument is suggestive rather than decisive. That candidates for consciousness are all electrochemical is consistent with electrochemistry being necessary — and equally consistent with electrochemistry being the only route evolution found to the computational complexity that actually matters. Block's own framing concedes that the advantage may be functional.

So this is not a refutation of Parts One and Two. It is an argument that their conclusion is conditional on a premise which is currently a bet rather than a finding — and it happens to be the premise that determines the answer.

5.Where the Series Lands

A method exists for assessing AI systems for consciousness that examines architecture rather than behaviour, draws indicators from several theories at once, and treats them as shifting probability rather than delivering verdicts. That is a real advance over argument by assertion.

Applied to real systems, it returned no current candidates, and no obvious technical barrier to building one that satisfies the list.

And the whole apparatus is conditional on computational functionalism, which excludes by construction the major theory that disagrees, and which a serious philosopher argues has not been established.

The honest summary is that the field has built a good instrument for a question it cannot yet confirm the instrument measures. That is not a failure. It is what a science looks like before its foundations are settled — and it is considerably better than what this debate consisted of five years ago.

One gap in our own coverage should be stated. The exchange continued after these papers: a critique of the indicator method by Cyriel Pennartz, and a reply from Butlin, Bayne and Fleming, both in the same journal. Both are behind a paywall with no obtainable copy, so we have not read them and this series does not reflect them.

Editorial Comment

MindHeaven® sells supplements and has no stake in whether machines can be conscious. We wrote three parts on it because the reasoning is unusually clean and the stakes are unusually honest on all sides.

The transferable lesson is Block's fork. A criterion that seems to be about evidence can turn out to be a choice made before the evidence, which then determines what the evidence can show. Decide to judge a compound by its mechanism and you will find promise; decide to judge it by outcomes in people and you will often find much less. Both are defensible; the mistake is not noticing you chose.

We would rather flag that in a domain where we have nothing to gain, so that a reader can hold us to it in the domain where we do.

Nothing in this series suggests that any system now in existence is conscious, and nothing in it should be used to argue that one is.