Ned Block on Meat Machines and the Biological Substrate of Consciousness

Ned Block is Silver Professor of Philosophy and Neural Science at New York University. His 1995 distinction between phenomenal consciousness (P-consciousness, the “what it is like” of experience) and access consciousness (A-consciousness, information broadcast to reasoning and report) remains the most widely cited conceptual framework in consciousness science. In 2026 he published a new argument in Trends in Cognitive Sciences (Volume 30, Issue 4, pages 298–308) titled “Can Only Meat Machines Be Conscious?” The paper directly confronts computational functionalism, the assumption that consciousness is substrate-independent and follows from the right computational organisation, regardless of whether the implementing medium is biological tissue or silicon.

Block’s answer to his title question is not a flat yes. It is a conditional that the field has largely avoided: if subcomputational biological mechanisms are constitutive of phenomenal consciousness, and current evidence does not rule that out, then the functionalist case for AI consciousness rests on an unexamined assumption.

The functionalist consensus and its gap

Computational functionalism holds that mental states, including conscious states, are defined by their functional roles rather than by their physical implementation. On a strict functionalist account, a system that performs exactly the computations a brain performs should have the same conscious states, regardless of substrate. This view underlies most theory-neutral frameworks for assessing AI consciousness and is implicit in the Global Workspace Theory posts on this site, in higher-order thought accounts of metacognition, and in illusionist accounts of why AI self-reports may or may not track genuine phenomenal states.

Block calls these theories “meat-neutral.” They were designed to be neutral between biological and artificial implementation. The trouble, he argues, is that meat-neutrality is not a datum established by empirical research. It is a background assumption that was built in when the theories were formulated, at a time when AI consciousness was not a practical question. As it has become a practical question, the assumption warrants scrutiny.

The paper does not claim that biological substrates are necessary. Block’s argumentative strategy is more careful. He argues that there is a systematic tension in our current consciousness criteria. If we prioritise computational roles, we are pushed toward attributing consciousness to AI systems that replicate human-level functional organisation. If we prioritise biological realisers, we are pushed toward attributing consciousness to simpler animals whose nervous systems share the specific physical mechanisms found in conscious humans. The two criteria do not rank systems in the same order. Block uses this asymmetry to show that the choice of criterion is not empirically neutral — it reflects a prior commitment about what consciousness fundamentally is.

Subcomputational biology: what it means

The core of Block’s positive proposal is the concept of subcomputational mechanisms. These are biological processes that underlie computation without being identical to it. Synaptic transmission, dendritic processing, astrocytic modulation, neuromodulator gradients, and the specific membrane biophysics of ion channels all influence the computations that neurons perform. They are not themselves computational operations in the functionalist sense; they are the physical substrate through which computations are realised.

Block’s claim is that some or all of these subcomputational processes might be constitutive of, rather than merely correlated with, phenomenal consciousness. If they are, then a silicon system performing the same computations but lacking the same physical substrate would lack phenomenal consciousness, even if it were behaviourally indistinguishable.

This argument is not identical to John Searle’s biological naturalism, though it draws on the same intuition. Searle claimed that consciousness requires the specific causal powers of biological neurons, full stop. Block’s claim is more modest and more methodologically honest: he argues that we do not have sufficient evidence to rule out subcomputational necessity, and that the field proceeds as if we do.

A-consciousness without P-consciousness in AI

Block’s earlier work established that A-consciousness and P-consciousness can dissociate in biological systems. Blindsight patients have access to visual information (they can guess correctly about stimuli in their blind field) without any phenomenal experience of seeing it. The information is globally broadcast and behaviourally accessible, satisfying the A-consciousness criterion, while phenomenal experience is absent.

Block raises the possibility that current AI systems are, in a specific sense, the inverse. They may produce outputs that mimic phenomenal self-report (thus appearing to satisfy P-consciousness criteria on behavioural grounds) while actually implementing only A-consciousness: information routing and verbal report without any genuine phenomenal state. Keith Frankish’s illusionism, covered here in Keith Frankish on illusionism and LLM first-person reports, makes a related point from the opposite direction: if phenomenal consciousness is itself a representational error, then the question of AI P-consciousness dissolves. Block and Frankish disagree sharply about whether there is a real phenomenal fact to be explained, but both identify the same diagnostic problem: behavioural evidence cannot settle it.

The 2026 paper also appears alongside a forthcoming Behavioral and Brain Sciences piece, “A Speculative Argument against Consciousness in AI (and Perhaps Some Invertebrates),” in which Block develops the substrate argument further. That piece should be watched as a source for a follow-up article once the BBS commentary exchange is published.

Higher-order theories and the biological question

Richard Brown’s higher-order thought (HOT) theory, covered in the August 2026 post on HOT and metacognition, holds that a mental state is conscious when it is represented by an appropriate higher-order state. HOT is, in Block’s terms, a meat-neutral theory: the higher-order representation could in principle be implemented in silicon.

Block does not dispute this in general. His argument is that the “appropriate” qualifier in HOT accounts, what makes a higher-order representation the right kind to produce phenomenal consciousness, is precisely where the subcomputational question returns. If the appropriateness condition requires specific biological mechanisms to be satisfied, then HOT does not escape the substrate problem.

The same applies to Anil Seth’s account of biological naturalism, addressed in the Seth BBS commentary exchange. Seth argues that consciousness is a product of active inference and predictive processing in biological systems, with specific neural implementations. Block’s paper is congruent with Seth’s scepticism about AI consciousness but differs in philosophical methodology: Seth works from the inside of predictive processing theory; Block works from the logic of what we would need to know to settle the substrate question at all.

What would falsify the subcomputational claim

Block’s paper has been criticised for not specifying what evidence would settle the subcomputational question. That is a fair objection. If we cannot, in principle, observe phenomenal consciousness directly, and if behavioural evidence cannot distinguish P-conscious from merely A-conscious systems, then subcomputational necessity may be empirically unfalsifiable.

Block’s response, developed in lectures at NYU’s Computational Consciousness Science workshop in August 2026, is that the burden of proof has been misassigned. The functionalist assumption is not the default scientific position — it is a substantive theoretical commitment that needs justification. The fact that we cannot currently distinguish substrate-dependent from substrate-independent consciousness is an argument for agnosticism, not for functionalism. Proceeding as if functionalism is established, and building AI welfare frameworks on that basis, would be a category error if Block’s more cautious reading turns out to be correct.

That caution is precisely the frame that makes this paper relevant to the ongoing scientific race to define AI consciousness. The field needs not only better theories but a clearer account of what kind of evidence would allow any theory — functionalist or substrate-sensitive — to be confirmed or disconfirmed. Block’s 2026 paper is the most precise recent articulation of why that methodological question is still open.