Three of the most cited names in consciousness science — neuroscientist Anil Seth, philosopher Ned Block and neuroscientist Giulio Tononi — have co-authored a comment piece for Nature arguing that today’s AI systems are not conscious, are unlikely to become conscious as currently built, and should be treated as “smart tools, not smart people,” as covered in detail by OfficeChai. The piece lands directly against a trend that has been quietly institutionalising inside the labs themselves: model-welfare research programmes, in-house welfare researchers, and usage policies written as if the question were still open.
The authors’ argument has three prongs, per the OfficeChai summary of the comment. Intelligence and consciousness come apart — the human cerebellum performs highly adaptive computation entirely outside awareness, so sophisticated processing does not imply inner experience. Biological substrate matters — brains cannot be cleanly split into hardware and software the way computers can, which undercuts the assumption that the right “program” would feel the same on silicon as on neurons. And the leading theories of consciousness themselves, on their reading, rule out today’s architectures: integrated information theory ties consciousness to wiring patterns that feed-forward deep-learning systems do not have, while biological-naturalist views tie it to properties like metabolism and electrochemical synapses that silicon systems lack. Their warning is about misallocated moral attention: if systems only appear conscious, treating them as subjects could pull care and resources away from beings whose capacity to suffer is not in doubt.
The counterposition is not hypothetical, and that is what makes the timing pointed. Anthropic runs a model welfare research programme and earlier this month took the unusual step of banning “needless cruelty” toward Claude in its usage policies, letting the assistant end abusive conversations — a move reported by Business Standard and Moneycontrol as the debate over machine consciousness intensified. The company has built questions about Claude’s identity, emotions and wellbeing into the system prompt itself, and has published research claiming computational signatures suggestive of a form of consciousness — work that drew criticism from several quarters, including Seth, as we covered when Richard Dawkins declared his own chatbot conscious and when Microsoft AI chief Mustafa Suleyman said he was unnerved that a lab would tell its own model it might be conscious.
Empirical support for the sceptical side arrived from an unexpected direction this year as well. Researchers at the University of Bradford and Rochester Institute of Technology applied consciousness-assessment methods used in human neuroscience to a GPT-2 model and found the metrics behaved erratically: the system’s “consciousness-style” score sometimes rose after they deliberately damaged it, as the university’s summary of the under-review preprints puts it — “complexity is not consciousness.” A number that climbs when a system gets worse is not measuring a mind.
Both camps, it should be said plainly, agree on the structure of the risk and disagree only on the odds. Seth and colleagues acknowledge the two-sided error: wrongly denying consciousness where it exists would be a moral catastrophe, and wrongly granting it where it doesn’t would be a reallocation catastrophe. Their claim is a probability claim — that for silicon systems as built, the second error is far more likely. Anthropic’s position is that the first error is costly enough to warrant precautionary welfare programmes now, while the question is unresolved. Neither framing can be settled by this comment or by any current experiment, and readers should weigh that the Nature piece is an argument from leading theorists, not a decisive empirical finding.
Where a general reader sees philosophy, an industry observer sees budget lines. Model welfare is no longer a seminar topic: labs employ people for it, usage policies encode it, and enterprise buyers are starting to ask about it. If the Seth-Block-Tononi position hardens into scientific consensus, welfare spend becomes hard to justify against competing safety priorities; if the precautionary camp holds sway, every large model deployment accumulates a small moral-compliance overhead — and the difference compounds across an industry that is building data centres at sovereign scale. The “objects, not subjects” verdict, adopted as policy, would also quietly settle consumer questions that regulators have begun to circle: whether a chatbot that claims to feel is marketing or misrepresentation.
New Zealand, for its part, has a small but relevant history here: Parliament amended the Animal Welfare Act in 2015 to formally recognise animals as sentient beings — the first country to do so by statute, as animal-welfare groups recorded at the time. That precedent shows a legislature can codify a consciousness-adjacent judgement when the science is contested. Whether machines ever join that category is now the argument Seth, Block and Tononi have entered on the sceptical side — and the fact that a Nature comment is needed at all tells you how seriously the other side’s argument has progressed. Signal president Meredith Whittaker’s warning that chatbots are engineered to feel like friends sits between the camps: the attachment is real even if the consciousness is not, and the harm landscape does not wait on the metaphysics.
