Google DeepMind chair and Alphabet chief scientist Demis Hassabis on the human brain: “a biological approximation to a Turing machine”

Demis Hassabis has built his career on treating the human brain as the one working blueprint for general intelligence. This week a fragment of his became the emblem of everything a critic believes the AI industry gets wrong about minds, and the full sentence it came from is narrower, and more revealing, than the piece that travelled.

“a biological approximation to a Turing machine.”

The phrase anchors an essay by Benjamin Riley in The Verge, “Our minds aren’t equipped to handle AI.” Riley opens with the cybernetics pioneer Norbert Wiener, “The thought of every age is reflected in its technique,” and offers Hassabis, now chair of Google DeepMind and Alphabet’s chief scientist, as the clearest case of an industry that sees its own machines when it looks at a skull. The words themselves come from a conversation with James Manyika published this year in Dædalus, the journal of the American Academy of Arts and Sciences.

What the full sentence says

In Dædalus, Hassabis was explaining what AGI means to him. He defines it as “a system that exhibits all the cognitive capabilities the human mind has,” because the brain is “the only existence proof that we know of so far, perhaps in the universe, that general intelligence is even possible.” Then comes the line in full: “The brain can be viewed as a biological approximation to a Turing machine, meaning that, in theory, it could learn anything that is computable.” The next sentence applies the brake. Any finite system needs “some degree of specialization because there is only so much time, memory, and information.”

Read that way, it is not a claim that people are laptops. It is a claim about range, about what a brain could learn in principle, and it doubles as an engineering spec. If the brain is the reference machine, AGI is whatever matches its entire range. Hassabis even sketches the exam: train a model with a knowledge cutoff of 1910 and see whether it arrives at General Relativity, as Einstein did in 1915. “For now, the answer is clearly no,” he says.

The critique works one layer down

Riley’s objection sits somewhere else. Drawing on Paul Cisek, a neuroscientist at the University of Montreal, he argues that brains are better understood as feedback-control systems that act on the world to change what they perceive, shaped first by evolution and then by culture: imitation, language, writing, schools. On that model, a chatbot that does the effortful part of thinking is “a cognitive hot dog,” tempting in the moment and corrosive as a regular diet. He points to OpenAI’s education chief Leah Belsky calling ChatGPT “the world’s largest learning platform,” and to a study from China in which thousands of students essentially stopped doing their homework once they started using AI.

The clause that decides it

The sharpest collision is inside Hassabis’s own paragraph. A few lines on, he says the key is “the brain’s potential to learn almost anything, especially as we are able to design and build tools and machines to help us gain knowledge.” That clause carries the whole dispute. Riley is not really proving the Turing-machine framing false. His evidence is about tools that do the learning instead of helping a person gain it. A computability claim describes what a brain could absorb in theory. It says nothing about whether a given product leaves the learner’s head fuller or emptier.

That question now lands on the person setting Alphabet’s scientific direction. Hassabis gave up the day-to-day CEO job in August to concentrate on AGI strategy and science, telling staff, as Fortune reported, that AGI is close at hand and that getting the next steps right is critical for humanity. A lab that measures its machines against the full range of the human brain has a direct stake in that range not shrinking in the people who use them.

An approximation to a Turing machine still has to be programmed by something. For people, that has always been the slow work of learning, and it is the one step a tool can either amplify or quietly skip.

Tags: Demis Hassabis, Google DeepMind, AI in education, Artificial general intelligence, Turing machine