Nobel laureate Demis Hassabis says the race to build AI that improves itself is outpacing the safeguards around it — and he is comparing the likely failure mode to an oil spill. The DeepMind co-founder, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, laid out his case on the latest episode of the Masters of Scale podcast. His core claim is simple: the technique works brilliantly, and that is precisely the problem.

The technique in question is recursive self-improvement, or RSI — AI systems that help produce better versions of themselves. Hassabis called it "an extremely useful technique" that nonetheless "carries the most risk." Note the pairing. He is not dismissing the approach; he is saying its power and its danger scale together. That distinction got less attention than the oil spill metaphor, and it matters more.

How the machine teaches itself

DeepMind demonstrated the underlying force years ago with AlphaZero. The system learned chess by playing millions of games against itself, starting from random moves. Hassabis recounted the timeline with some disbelief of his own: the system was still beatable at 11 a.m., reached grandmaster level by lunch, and "by teatime it was better than any chess-playing entity, human or computer." A single afternoon took it from novice to beyond every human and machine that came before.

That trajectory was possible, Hassabis explained, because of one property: verifiability. RSI works best in domains where you can check the answer. "Is the equation correct? Does the code run? Did you win the game?" Games qualify. Mathematics qualifies. And now code qualifies too — which, he argued, is a large part of why AI coding tools have improved as fast as they have. Where success can be measured automatically, self-improvement compounds quickly.

Why the next step is the dangerous one

Here the story turns on a number and a warning. AI is currently making engineers three to five times more productive, Hassabis said — with humans still in the loop, reviewing and writing code. That multiplier is the headline the industry prefers. The caveat follows immediately: the productivity gains exist only because a person is still checking the output.

The next step, taking the human reviewer out, is what Hassabis called "a very dangerous way" to get to smarter systems. The logic is his own and worth stating plainly. If AI now writes much of the code that trains the next AI, and you remove the human who verifies that code, you get improvement without oversight. Errors, or worse, compound through the loop the way a spill spreads through water. He did not, on the evidence available, specify when fully autonomous self-improvement might arrive or what regulatory mechanism should restrain it. The warning is directional, not operational.

What does this change for the industry?

Practically, two things. First, Hassabis — a co-founder of the laboratory most advanced in reinforcement learning (training systems by reward and penalty rather than by copying human examples) — is confirming that self-improvement is now a central engineering path, not a theoretical one. The AlphaZero anecdote is the proof of concept; the coding-tool boom is the proof at scale. Second, he has put a marker down on where the red line sits: humans in the review loop. Companies racing to automate that review are, by his account, racing past the safeguard.

What he did not say is also part of the record. He offered no timeline, no assessment of whether current safety teams can audit self-improving systems at the speed they now iterate, and no view on whether rivals beyond DeepMind are already testing human-free loops. Each of those gaps is where an oil spill would start. The metaphor implies a contained, visible, eventually cleanable disaster; whether an error propagating through recursive code improvement stays containable is exactly the question Hassabis left open.

For executives, the near-term reading is sober but actionable. The three-to-five-fold productivity figure is real and current, and it depends on human reviewers. Any plan that budgets for removing those reviewers should account for the risk the field's most decorated researcher has now named twice in one interview. The industry's course, in his framing, is still changeable — but only while someone is still reading the code.