Geoffrey Hinton puts the odds of human extinction at 50%. Should we believe him?

Key points

• Geoffrey Hinton, one of the founding researchers of modern AI, estimates a 50% probability of human extinction caused by AI.
• Elon Musk predicts AI will bring "an age of amazing abundance" but puts a 10 to 20% chance on AI systems killing humanity.
• Researchers in Silicon Valley believe recursive self-improvement, the process where AI trains the next version of itself, could arrive within years.
• No expert quoted here agrees on whether the outcome will be good or catastrophic, only that something large is coming fast.

Geoffrey Hinton is not a fringe voice. He shared the Nobel Prize in Physics in 2024 for work that made modern AI possible. When he says something scary, read it twice.

His estimate of p(doom), shorthand researchers use for the probability that AI ends human civilisation, sits at 50%. Not because he has calculated it carefully. "Because I haven't got a clue how to estimate the real number," he told Sebastian Mallaby, author of The Infinity Machine. A coin flip, because the honest answer is that nobody knows.

What is actually scaring these people?

The specific fear is recursive self-improvement. Today, humans design each new AI model. Recursive self-improvement is the moment AI starts designing the next AI itself, producing a better version, which produces a better one still, with no human hand on the wheel.

Experts in Silicon Valley, as The Guardian reported, believe this moment could arrive within years. One Stanford researcher greeted a colleague with: "Welcome to the foothills of the singularity," meaning the climb has already started. We have been tracking this question since our first recursive self-improvement story on 29 July 2026, when a startup founded by former Anthropic researchers signed a $100 million-plus deal with Google to build systems designed to get smarter on their own.

Hinton's view is blunt. "As soon as evolution kicks in, we're fucked," he said. He said it cheerfully, apparently.

Is anyone optimistic?

Elon Musk is, sort of. He predicts "an age of amazing abundance" once superintelligent AI arrives, while simultaneously placing a 10 to 20% chance on killer robots murdering us all. That is a strange thing to hold in the same breath.

Author Robert Wright, in The God Test, frames it more carefully: never before has the near-term future held such a wide array of paths for humanity that would be so wildly consequential.

The people closest to the technology disagree violently on the outcome. They largely agree the technology is coming regardless.

What does this mean for ordinary people?

Nothing actionable this week, and that is partly the problem. AI is already reshaping jobs and creative work well before any superintelligence arrives. The OpenAI safety team dissolution we reported on 16 August is the kind of institutional signal worth watching: the people paid to spot danger inside the most powerful lab in the world no longer have a team.

Understanding what AI can and cannot do today gives you more ground to stand on when the next leap happens. This is not a passive debate to observe from a distance.

Common questions

What is superintelligence, exactly?

Superintelligence means an AI system that outperforms humans across most intellectual tasks, not just chess or image recognition but judgment and reasoning. No such system exists today, though researchers disagree sharply on how far away one is.

Should I be worried right now?

The risks experts discuss are real but not immediate in the extinction sense. What is immediate is that today's AI already affects jobs, media and decision-making, so understanding it is practical, not paranoid.

Why can't experts agree on the risk?

Because recursive self-improvement has never happened before. There is no data to run the numbers on. Hinton's 50% figure is his way of saying the honest answer is genuine uncertainty, not false confidence in either direction.