In the October 22 issue of Nature, Anil Seth, Ned Block, and Giulio Tononi (Tononi from my old UW–Madison home) pool their decades of work in experimental neuroscience, theoretical neurobiology, and philosophy of mind into a pointed Comment. They argue that current AI systems are not conscious and that, on present trajectories, they are unlikely to become so. The piece is partly a response to recent claims from figures such as Richard Dawkins and Geoffrey Hinton, and to Anthropic’s 2026 preprint reporting computational signatures of a form of consciousness in its Claude model.
The authors frame the stakes in practical terms. If AI systems really are conscious, they might suffer, and proliferating them raises serious moral questions. If they only seem conscious, then granting them rights or welfare protections could hobble oversight just when it is most needed, and could divert concern from beings whose sentience is not in doubt. They reject a precautionary stance: uncertainty does not make all possibilities equally likely, and in their judgment false positives (attributing consciousness where there is none) are far more likely than false negatives.
They offer three reasons for skepticism.
1. Intelligence is not consciousness. Consciousness is what disappears in dreamless sleep or under anesthesia, what Nagel described as there being “something that it is like” to be an organism. Intelligence concerns doing, while consciousness concerns being and feeling. The cerebellum makes the point. It holds roughly four-fifths of the brain’s neurons and does sophisticated, adaptive motor computation, all outside awareness, and losing it does not directly affect consciousness. Our tendency to merge the two reflects anthropocentric and anthropomorphic biases, amplified by the seductive power of language. LLMs, trained on vast amounts of human writing about inner life, act as mirrors of our collective digitized past. As the authors put it, “Just because something is intelligent does not mean that it is conscious.” The contrast case is AlphaFold. It is also a powerful neural network on silicon, but because it predicts protein structures instead of chatting with us, almost no one wonders whether it is conscious.
2. Brains are not computers made of meat. The case for conscious AI rests on computational functionalism, the assumption that the right computation suffices regardless of substrate. The authors trace this idea to Turing’s 1936 universal machine and to McCulloch and Pitts’ 1943 logical neurons. They argue the brain-as-computer metaphor has been overextended. In biological brains, what the system does cannot be cleanly separated from what it is. There is no software/hardware divide in living neural tissue. Evidence that substrate matters is therefore evidence against functionalism, and against consciousness in silicon.
3. Leading theories of consciousness are non-computational.
- Integrated information theory (IIT) holds that consciousness requires a physical system whose components are interconnected in specific ways that satisfy the essential properties of experience. Cortical regions meet these requirements. The cerebellum does not, and neither do the hardware architectures that run deep learning, however capable those systems are.
- Biological naturalism proposes that consciousness depends on particular material features of living systems. Candidates include electrochemical synapses, cortex-wide electromagnetic fields, metabolism, and autopoiesis (self-regeneration).
Their conclusion is that the research community’s default position should be that silicon-based AI is not conscious and should not be treated as such. Such systems are, in their phrase, objects rather than subjects, “smart tools, not smart people.”
Readers of MindBlog will recognize this as one side of a debate I have followed for years. The predictive-processing and Agüera y Arcas perspectives I have often featured lean toward functionalism, and the authors’ substrate argument is the crux on which they part company.
No comments:
Post a Comment