Seth, Block, Tononi: AI not conscious, perhaps never will be

Nearly four-fifths of the neurons in the human brain sit in the cerebellum, and you cannot feel any of them working.

That is the empirical pivot Anil Seth, Ned Block and Giulio Tononi lean on in a new Nature comment arguing that today's AI is not conscious, and perhaps never will be. Seth runs the Sussex Centre for Consciousness Science; Block is at NYU philosophy; Tononi directs the Wisconsin Institute for Sleep and Consciousness. Three of the researchers we track shared the piece the day it ran.

Their first move is a definition. "Consciousness is about being and feeling. Intelligence, very broadly, is about doing, or performing functions," they write, borrowing Thomas Nagel's framing of experience as "something that it is like to be that organism." The cerebellum is the proof: it "adjusts and plans motor performance in a way that is precise, fast and adaptable through ongoing learning," yet its absence or removal "does not affect consciousness directly." Sophisticated computation and felt experience come apart inside a single brain, so a system doing clever things is not, by itself, evidence that anything is being felt.

Second move: computational functionalism, the view that running the right program on any substrate is enough for experience, is "fatally flawed." "Brains are not, or not just, computers made of meat," they write. "Computations clearly suffice for forms of intelligence...but it is much less clear that they could suffice for consciousness."

Third: the leading theories of consciousness do not fit the hardware. Integrated information theory, which Tononi developed, requires a system's components "being wired together in very specific, tightly interconnected ways" that the "layered, feed-forward" architecture of deep-learning systems does not have. Large language models, in their phrase, are "mirrors that reflect an image of our collective digitized past."

They name Anthropic directly. The company's recent research claiming to have identified "computational signatures of a form of consciousness in Claude" has, they write, "drawn criticism from several quarters, including from Seth himself." They also note Anthropic has embedded questions about Claude's "identity, emotions and wellbeing" into training documents.

Their bottom line is a policy line as much as a scientific one. Treat today's AI as "smart tools, not smart people." A precautionary stance on machine consciousness, they argue, risks diverting "attention and resources away from beings whose capacity to suffer is not in question."