Seth, Block and Tononi: AI isn't conscious, probably won't be

Current AI systems are not conscious, and along current trajectories of development, almost certainly never will be.

That is the argument in a Nature comment from neuroscientist Anil Seth of the University of Sussex, philosopher Ned Block of NYU, and neuroscientist Giulio Tononi of the University of Wisconsin–Madison, published on 19 October. The authors define consciousness as 'what goes away with dreamless sleep or under general anaesthesia': subjective experience, there being 'something that it is like to be that organism.' Today's large language models, they argue, are not candidates for it. Two researchers we track circulated the piece on the day it ran.

The first move separates intelligence from consciousness. 'Consciousness is about being and feeling; intelligence is all about doing, or performing functions,' the authors write. The human cerebellum holds 'nearly four-fifths of the neurons in the human brain' and performs 'precise, adaptive, highly intelligent motor computations entirely outside conscious awareness'; damage to it 'does not affect consciousness directly.' Language models, in their framing, 'are mirrors that reflect an image of our collective digitized past.' Fluent because trained on human-written text, not because anything is felt.

The second move rejects computational functionalism, the view that the right algorithm suffices on any substrate. 'Brains are not… computers that are made of meat,' the authors write. On the account they favour, which leans on integrated information theory, 'biological properties are necessary for consciousness,' with candidates including electrochemical synapses, structures that integrate neural activity across broad regions of the cortex, and 'metabolism and autopoiesis.'

From there the paper lands on a practical rule for the field: 'The default position of the research community should therefore be that AI is not conscious.' The authors argue that 'false positives are much more likely to occur than false negatives are,' and that a precautionary stance built on treating current systems as potentially sentient would be 'misguided.' Their verdict: today's AI are 'smart tools, not smart people.'