DeepMind's Hassabis Warns of an AI 'Oil Spill' Without New Safeguards

DeepMind co-founder Demis Hassabis says the race to build AI that improves itself is outpacing the safeguards around it — and he warns an AI "oil spill" is coming unless the industry changes course. He made the comments on the latest episode of the Masters of Scale podcast.

Hassabis is not a casual observer. He shared the 2024 Nobel Prize in Chemistry for AlphaFold, the protein-structure breakthrough, and co-founded DeepMind, the lab behind some of the field's defining systems. Few people, as the interview framing put it, have shaped modern AI more.

His target is a specific technique: recursive self-improvement, or RSI — systems that help produce better versions of themselves. AI is now building better AI. Hassabis says the next step, taking the human reviewer out of the loop, is "a very dangerous way" to get there.

What is recursive self-improvement?

RSI describes AI systems that generate improvements to themselves, compounding capability with each cycle. DeepMind demonstrated the technique's power years ago with AlphaZero, the system that learned chess by playing millions of games against itself.

The timeline Hassabis recounts is blunt. Starting from random moves, AlphaZero was still beatable at 11 a.m. It reached grandmaster level by lunch. "By teatime it was better than any chess-playing entity, human or computer," he said.

That quote carries weight because it's a measured observation, not hype. AlphaZero went from ignorant to superhuman inside a single working day.

Where does RSI work best?

Hassabis draws a clear boundary around the conditions that make self-improvement productive. RSI works best, he said, in domains where you can check the answer.

"Is the equation correct? Does the code run? Did you win the game?" he asked, listing the verification questions that make self-play and self-correction reliable.

Games, math, and now code all qualify, according to Hassabis. That verifiability is a large part of why AI coding tools have improved as fast as they have. When a machine can test its own output against a runnable program or a provable result, each iteration compounds.

Why is the next step the risky one?

Today's systems still have humans in the loop. Hassabis said AI is making engineers three to five times more productive — with people still reviewing and writing code alongside the machine.

That multiplier is the current state of the art: augmentation, not autonomy. The risk sits one step further out, when AI writes and reviews the improved AI without a human checking the work.

Hassabis doesn't dismiss the technique. He called self-improvement "an extremely useful technique" while insisting it "carries the most risk." Both statements come from the same person in the same conversation, and he offers them without contradiction — the usefulness is exactly why the danger compounds.

The "oil spill" framing makes his position concrete. Hassabis isn't predicting deliberate misuse. He's predicting an accident: a fast-moving, self-improving system that outruns the containment built around it, the way a rig failure outruns a containment boom.

What's the so-what?

The warning lands at a moment when coding is the fastest-moving commercial application of AI, and the verification advantage Hassabis describes is precisely what makes code the natural first arena for recursive self-improvement. His argument implies the industry's most productive category is also the one closest to removing the human reviewer — the step he calls "a very dangerous way" to advance. Whether labs slow down before or after the first spill is now the open question.