- Intelligence is doing; consciousness is feeling. Language models show only the former.
- Seth argues brains can't be separated into software and hardware, unlike computers.
- A December 2025 paper stakes a middle position: biological computation isn't separable either.
Anil Seth wants you to notice how easily you are fooled about AI consciousness. Not by a magician, but by your own mind, which is built to find faces in clouds and, he argues, minds inside chatbots.
He is not being cute about it.
Why AI is unlikely to become conscious is the case Seth brought to the TED stage in Vancouver in April, 2026. He is professor of cognitive and computational neuroscience at the University of Sussex, where he directs the Sussex Centre for Consciousness Science, and his target is the leading language models, Claude and GPT included.
Faces in the clouds
The case starts with a distinction few of us bother to draw. Intelligence is doing: solving a puzzle, assembling furniture, navigating a tricky family dinner. Consciousness is feeling: the taste of coffee, the warmth of a fire, the fact that anything is happening at all inside.
They rarely go together outside our own heads.
Nobody worries whether DeepMind's AlphaFold is conscious, even though under the hood it is not much different from Claude or GPT: algorithms running on silicon, trained on vast reservoirs of data. AlphaFold predicts protein shapes rather than sentences, and sentences are what pull our psychological strings.
That asymmetry, Seth says, tells us more about human psychology than about artificial intelligence.
Do you see a face in the cloud? (Science Reader)
Not a computer made of meat
His deeper claim is about metaphor. The brain has been a plumbing system, then a telephone exchange, then a computer, and each metaphor was useful right up until it was mistaken for the truth.
The brain is not, or at least not just, a computer made of meat.
Anil Seth, University of Sussex
Unlike a laptop, brains keep no clean line between software and hardware. Neurotransmitters wash through the tissue, electromagnetic fields sweep the cortex like weather, and a single neuron is a stranger, richer machine than the cartoon units that power today's AI. What a brain does, on this account, cannot be peeled away from what a brain is.
Two camps in the AI consciousness debate
Computational functionalism holds that the right pattern of information processing is enough to produce consciousness, whatever hardware runs it. Biological naturalism, Seth's position, insists that living tissue itself is doing indispensable work, not just hosting a program.
A hurricane simulated in software does not generate real wind, Seth points out, and a brain simulated in enough detail would not necessarily generate real experience either. If he is right, consciousness will need biology, not just code, and today's AI is nothing but code.
So far, this is a tidy argument.
Where the AI consciousness debate goes next
It is also not the only position on the table. In December 2025, the neuroscientists Borjan Milinkovic, of the Paris-Saclay Institute of Neuroscience, and Jaan Aru, of the University of Tartu, published a paper making a related but distinct case. In biological systems, they argue, the algorithm is the substrate, not something separable from it.
Their conclusion lands close to Seth's without being identical: they are less interested in banning computation from the story than in insisting that no clean layer of software ever floats free of its biological hardware.
Key figure
20%
How often Claude Opus 4 and 4.1 named an artificially injected concept, at the best injection layer and strength, with close to zero false positives
Two months earlier, evidence of a different kind had arrived from inside the machines themselves. Jack Lindsey, a researcher at Anthropic, found that Claude's Opus models could sometimes detect when a concept had been artificially injected into their own activations and name it correctly.
Lindsey is careful to note that detecting an injected pattern is not the same as feeling one. Seth would put it more bluntly: a thermostat that reports its own temperature is not thereby cold or warm to itself.
Philosophers on the functionalist side disagree; if a system's internal organization does the right causal work, they argue, the biological question may not matter at all.
Seth thinks the stakes go beyond philosophy seminars. If AI merely seems conscious, without being conscious, granting it rights would hand away our ability to regulate or switch it off for no good reason. And a machine that seems to feel for us is a machine we may trust past the point sense would allow.
Let's hope the machines aren't looking for faces in the clouds on their own quite yet.
Sources
- Primary Source: Why AI is unlikely to become conscious (Anil Seth, TED)
- Additional Context:
- On biological and artificial consciousness: A case for biological computationalism (Milinkovic & Aru, Neuroscience & Biobehavioral Reviews, 2025)
- Emergent Introspective Awareness in Large Language Models (Jack Lindsey, Anthropic, 2025)
- AI, Consciousness and Ethics symposium (AISB Convention 2026, University of Sussex)
Fact Check: Claim-by-Claim Verification Verified
Seth's argument is reported accurately and every quotation and paraphrase from the talk was checked verbatim against TED's official transcript, with no misquotation found. Two factual errors were corrected: the Anthropic introspection paper was published two months before the Milinkovic and Aru paper, not in the same month; and a clause describing today's AI as "pure physical machinery" inverted Seth's actual claim that AI is software. Two overstatements were also tightened.
Editor note: wording left unchanged, since "unlikely" is Seth's own and is more faithful to what he argues than TED's flattened title.
Editor note: the draft said AlphaFold "runs on the same basic architecture as Claude or GPT." That overstates both Seth and the engineering, since AlphaFold uses specialized modules rather than a language-model transformer stack. Rewritten to Seth's own weaker and accurate formulation, "under the hood it is not much different."
Editor note: "In December" made explicit as "In December 2025," since the sentence now anchors a date comparison.
Editor note: corrected to "Two months earlier, evidence of a different kind had arrived."
Editor note: the key-figure label now names the condition.
Editor note: rewritten to say Seth keynoted the convention around the symposium and that one session was set aside for biology against substrate independence.
Commentary
- Every quotation and close paraphrase from the talk was checked against TED's official transcript, retrieved from the embedded page data rather than a summary. The transcript is 2,241 words across 33 paragraphs and matches the copy verified for the earlier Science Reader piece on the same talk.
- The talk was retitled after publication. TED now calls it "Why AI isn't going to become conscious," which is firmer than Seth's own "unlikely." Readers clicking the source link will land on the newer title.
- Whether consciousness requires biology is unresolved and the article does not pretend otherwise. Biological naturalism is a minority position under active dispute, not a settled finding.
- The 20 percent introspection figure is a best-case result at an optimized injection layer and strength. On most trials the models fail, which the paper states plainly.
- Perplexity's search pass introduced two errors that primary sources overturned: it identified the Santiago meeting as ASSC 30 (it was ASSC 29; ASSC 30 is Beijing 2027) and it misspelled Milinkovic's first name as Bojan. Both were corrected against the societies' and PubMed's own records.
Sources used for verification
Academic/Peer-reviewed:
- On biological and artificial consciousness: A case for biological computationalism - Neuroscience and Biobehavioral Reviews (Milinkovic and Aru)
- PubMed record 41419099, with author affiliations and online date - pubmed.ncbi.nlm.nih.gov
- Adversarial testing of global neuronal workspace and integrated information theories of consciousness - Nature
- Functionalism - Stanford Encyclopedia of Philosophy
Other reliable sources:
- Official transcript, Anil Seth at TED2026 - ted.com
- Emergent Introspective Awareness in Large Language Models - transformer-circuits.pub
- Past conferences, including ASSC 29 in Santiago - theassc.org
- AISB Convention 2026, Sussex, with Anil Seth as keynote - aisb.org.uk
- AICE-26 Symposium on AI, Consciousness and Ethics programme - aice-symposium.github.io
- CogSci 2026, Rio de Janeiro, 22 to 25 July 2026 - cognitivesciencesociety.org
- Anil Seth staff profile - sussex.ac.uk
- TED2026 conference details - conferences.ted.com
Fact-checked by Perplexity Sonar Pro on 2026-08-16
