nature.com web signal

Seth, Block, Tononi: AI isn't conscious, may never be

TL;DR

  • Three leading consciousness scientists argue in Nature that today's AI isn't conscious and perhaps never will be, pushing back on Hinton and Dawkins.
  • They note the cerebellum contains nearly four-fifths of the brain's neurons but its removal doesn't directly affect consciousness: cognition and experience can diverge.
  • Integrated information theory and biological naturalism, they argue, both predict current feed-forward deep-learning architectures and silicon substrates lack what consciousness requires.

Three of the most cited researchers on consciousness, writing together in Nature, say today's AI systems are not conscious and perhaps never will be.

The 19 October 2026 piece in Nature is by Anil Seth, a cognitive and computational neuroscientist at the University of Sussex; Ned Block, a philosopher at New York University; and Giulio Tononi, who directs the Wisconsin Institute for Sleep and Consciousness. They offer three arguments against the view, voiced recently by Geoffrey Hinton and Richard Dawkins, that large language models might already have inner lives.

First, they prise intelligence apart from experience. "Consciousness is about being and feeling; intelligence — very broadly — is all about doing, or performing functions," they write. They point to the cerebellum, which "contains nearly four-fifths of the neurons in the human brain," and note that its removal does not directly affect consciousness. A great deal of cognitive work can happen with the lights off.

The second argument targets a background assumption of much AI-sentience talk: that if you run the right program on any substrate, experience must follow. The authors reject this computational functionalism in one blunt line. "Brains are not — or at least not just — computers that are made of meat."

The third reason reaches for named theories. Tononi's integrated information theory treats consciousness as a product of densely interconnected causal wiring, which the authors say current feed-forward deep-learning systems lack. Biological naturalism, the other theory they invoke, ties experience to features such as electrochemical synapses and metabolism that silicon does not share. On either theory, today's models fail at the architecture or at the substrate.

The piece engages directly with reports from Anthropic claiming to identify consciousness signatures in Claude, treating such claims as a case study in anthropomorphism: a system trained on vast amounts of human text speaks like a human, and human readers oblige by reading a mind into it.

Shared on Bluesky by 5 AI experts