Approximate Turing Machine
In one sentence
The approximate Turing machine hypothesis holds that both biological brains and modern AI systems are best understood as imperfect, probabilistic implementations of the theoretical universal computer — with the consequence that what AI can ultimately do is an empirical question, not a philosophical one.
The Turing machine background
Alan Turing’s 1936 paper described a theoretical device — the Turing machine — capable of computing anything that could be expressed as an algorithm. Not any specific calculation, but any calculation: anything computable at all. Modern computers are, in the formal sense, Turing machines with memory constraints and speed limits. They can compute anything; they just cannot compute everything simultaneously or instantly.
The question that neuroscientists, cognitive scientists, and AI researchers have been arguing about for decades is whether the human brain is also a Turing machine — and therefore whether sufficiently capable AI might eventually match human cognition in any domain, or whether the brain does something that Turing machines cannot.
Hassabis’s position
Demis Hassabis stated it plainly: “a lot of neuroscientists including me think that maybe the brain… is an approximate Turing machine.” The qualifier approximate does the load-bearing work. A perfect Turing machine is lossless. Brains are noisy, lossy, probabilistic, constrained by metabolism and wiring. Modern AI systems are also noisy, lossy, and probabilistic — they produce different outputs from identical inputs depending on temperature settings and random seeds. Both are approximations of the theoretical ideal. Both are, in this framing, running the same kind of computation.
If the hypothesis is correct, the limit of what AI can do is not set by some deep qualitative difference between silicon and neurons. It is set by scale, training, and architectural ingenuity. Remove those constraints, and the question of what falls outside AI’s reach becomes genuinely open — which is a stranger and more vertiginous position than either the AI-will-never-be-conscious camp or the AI-will-surpass-us camp usually occupies.
What the hypothesis leaves open
The approximate Turing machine frame does not resolve the consciousness question. It brackets it. Whether a system that computes approximately as a Turing machine computes thereby has subjective experience — whether there is something it is like to be that system — is a question the frame deliberately does not answer. Hassabis was explicit about this: “I’m quite open-minded about what the answers might be.” That open-mindedness is itself a philosophical position: the avyākata move, the deliberate refusal to resolve a question before the evidence warrants it.
This Dictionary makes the same move, see Descartes Was Wrong. The approximate Turing machine entry names the computational structure without adjudicating whether computation of this kind is sufficient for mind or existence of a self. Both are useful framings. Neither is the last word.
The practical consequence
For the practitioner, the approximate Turing machine hypothesis has one immediate implication: the categories of work you believe AI cannot do are probably empirical claims, not necessary truths. They may be right today. They may not be right next year. The honest position is to hold them lightly, test them regularly, and update when the evidence changes. The history of AI is substantially a history of experts stating what AI can never do — and then watching it do that.
See also
Consciousness Calculator · Move 37 · The CERN Alternative · Sovereign Compute
Proposed May 9, 2026. Source: Demis Hassabis interview, Huge Conversations / Cleo Abram, May 2026; Alan Turing, “On Computable Numbers” (1936).