Three of the most cited names in consciousness science have come together to make a public case against machine sentience. In a new comment piece for Nature, neuroscientist Anil Seth, philosopher Ned Block and neuroscientist Giulio Tononi argue that today’s AI systems are not conscious, and are unlikely to become conscious even as they grow more capable.
The timing is notable. Questions about whether large language models might have some form of inner experience have moved from philosophy seminars into boardrooms and op-eds over the past year, with debate intensifying around Anthropic’s own public statements on the subject. The authors note that the question is no longer purely academic: if AI systems really were conscious, that would raise urgent questions about suffering, rights and welfare; if they only appear conscious, treating them as though they were could pull attention and resources away from beings whose capacity to suffer is not in doubt.
Why This Debate Is Heating Up
Seth, Block and Tononi point out that the idea of conscious AI is no longer a fringe position. They cite biologist Richard Dawkins and AI pioneer Geoffrey Hinton as prominent figures who have floated the possibility, and they note that Anthropic itself published research earlier this year claiming to have found computational signatures of a form of consciousness in Claude — research that has since drawn criticism from several quarters, including from Seth himself.
That broader unease has only grown as Anthropic has built questions about Claude’s “identity, emotions and wellbeing” into the document that governs how the model is trained, a move that has prompted even rival executives such as Microsoft AI CEO Mustafa Suleyman to say he’s nervous that a company would speculate about its own model’s consciousness inside the system meant to shape its behaviour.
Against this backdrop, the authors argue that uncertainty about AI consciousness does not mean all outcomes are equally likely. They say decisions in this area have to weigh the risk of wrongly denying consciousness where it exists against the risk of wrongly attributing it where it doesn’t — and that for today’s AI, the second kind of mistake is by far the more probable one. On that basis, they lay out three separate arguments for why they believe silicon-based AI, as currently built, is not a plausible candidate for subjective experience.
Argument One: Intelligence Is Not The Same Thing As Consciousness
The authors’ starting point is a basic distinction: consciousness is about what it feels like to be something, while intelligence is about what a system can do. Borrowing from philosopher Thomas Nagel’s famous framing, they describe consciousness as the quality of there being “something it is like” to be a given organism — the felt experience of tasting ice cream or seeing a sunset, as opposed to simply processing information about those things.
Seth, Block and Tononi argue that intelligent behaviour and conscious experience can and do come apart. The human cerebellum, they note, runs precise, adaptive, highly intelligent motor computations entirely outside conscious awareness, and a person can lose it without losing consciousness at all. The implication is that sophisticated information processing — the kind large language models clearly perform — does not by itself imply any inner experience.
The authors suggest that humans are primed to blur this distinction because of two built-in biases: anthropocentrism, or viewing the world through a human lens, and anthropomorphism, or projecting human traits onto things that merely resemble humans. Language, they argue, is an especially powerful trigger for this illusion, because fluent conversation is something humans have always associated with having a mind behind it. They contrast this with Google DeepMind’s AlphaFold, a similarly powerful neural network that predicts protein structures — a system almost nobody worries is conscious, precisely because it doesn’t talk back.
Argument Two: Brains Aren’t Just “Computers Made Of Meat”
The second argument takes aim at a foundational assumption behind most conscious-AI claims: computational functionalism, the idea that the right kind of computation is sufficient for consciousness regardless of what hardware runs it. The authors trace this idea back to Alan Turing’s 1936 proof that one sufficiently powerful computer can replicate any calculation performed by another, and to the 1943 work of Warren McCulloch and Walter Pitts, who showed that simplified networks of artificial neurons could carry out basic logical operations — laying the conceptual groundwork for both modern neural networks and the habit of treating the brain as just another information-processing device.
Seth, Block and Tononi reject this framing. They argue that biological brains cannot be cleanly separated into software and hardware the way digital computers can — that what a brain does and what a brain is made of are inseparable. If the biological substrate actually matters for how consciousness arises, they argue, that undercuts the Turing-style assumption that the same “program” would produce the same experience whether it runs on neurons or on silicon chips, making consciousness in today’s AI architectures far less likely.
Argument Three: Leading Theories Of Consciousness Rule Out Current AI
The final argument draws on specific scientific theories of consciousness. The authors highlight integrated information theory (IIT), which holds that consciousness depends on a system’s components being wired together in very specific, tightly interconnected ways — the kind of architecture found in parts of the human brain that support consciousness, but not in the cerebellum, and not, they argue, in the layered, feed-forward structure of today’s deep-learning systems. If IIT is correct, they write, no amount of raw capability can make current AI architectures conscious, because the wiring itself rules it out.
They also point to a family of theories grouped under “biological naturalism,” which hold that consciousness depends on specific material properties of living brains and bodies — things like electrochemical synapses, brain-wide electromagnetic fields, or basic biological processes such as metabolism and self-regeneration. None of these properties, almost by definition, exist in a digital system running on silicon.
The Stakes: Rights, Welfare And Where Attention Should Go
The authors are careful to frame their position as a rejection of a precautionary approach that has been gaining traction in some quarters — the idea that even a small chance of AI consciousness justifies extending some form of protection to AI systems now, in order to avoid repeating past moral failures such as the historical mistreatment of animals. Model welfare research of exactly this kind has become a small but fast-growing field over the past two years, with dedicated research groups and in-house welfare researchers now appearing at major AI labs.
Seth, Block and Tononi argue that this precautionary instinct is, in their view, misguided given how unlikely they believe AI consciousness to be. Their conclusion is blunt: today’s silicon-based AI systems should be treated as objects rather than subjects — “smart tools, not smart people,” as they put it — and that arguing otherwise risks diverting attention and resources away from the welfare of beings whose capacity to suffer is not in question.