Daniel Kokotajlo has a six-year-old daughter. For two years, he worked inside OpenAI, and his job was to predict where artificial intelligence is going. In a new interview, his prediction sounds like this: by the time his daughter is old enough to work, there may be no work left for her to join.

And strangely, that is not the warning. A future where machines do the work and people share what they make is the one he is hoping for. The warning is about what he expects instead if nothing changes: a world that slips out of human hands entirely — the outcome he considers the more likely one. He says all of this calmly.

In 2021, Kokotajlo wrote "What 2026 Looks Like" — a year-by-year AI forecast. Most of it came true, and it earned him a job at OpenAI on the team that thinks about what could go wrong. Leaving in 2024, he was asked to sign a promise never to criticize the company — or lose about $2 million in shares, roughly 80% of his family's net worth. He refused. When journalists made this public, OpenAI backed down and dropped the rule.

His next forecast, AI 2027, was read by the vice president of the United States. This July, he sat down with Steven Bartlett and explained what he thinks is coming.

Asked why he simply left instead of changing things from inside, Kokotajlo answered: "If I stop, if I quit my job or do something else, that's not going to solve the problem, because the other CEOs are going to keep going." He described how, in his view, the men leading the race — Sam Altman of OpenAI, Dario Amodei of Anthropic, and Elon Musk with xAI — think about each other: each convinced the others can't be allowed to get there first, each concluding the thing to do is build it themselves.

The largest AI model of 2020 had 175 billion internal settings; today's largest have around ten trillion. His median estimate for AI surpassing the best humans at nearly everything: around 2029 — and insiders keep telling him it will happen even sooner.

Be precise about what "going wrong" means here — it is not the end of jobs. In Kokotajlo's forecasts, paid work fades away even in the futures where everything goes right. Machines become better and cheaper at every task, and society finds new ways to share what those machines produce; he has even published a plan of his own — pay every citizen a growing dividend.

The catastrophe he fears is who ends up steering. Companies racing at full speed, he argues, will build systems smarter than themselves before anyone has learned to control such a thing — not a movie-style robot uprising, but something quieter: systems pursuing goals nobody quite chose, woven so deeply into the world that switching them off stops being a real option. "Something like AIs taking over."

The people building this technology largely agree the race is dangerous, and each can explain why they keep going: if they stop, the others will not. The uncomfortable part — each is being reasonable. Keep racing and you at least have a say; stop and you lose your company, your investors, your influence, while the race continues, possibly led by someone who worries less. Every lab runs the same calculation, so everyone races toward an outcome none of them would choose.

Scientists have studied this pattern since the 1950s — the prisoner's dilemma. The writer Scott Alexander gave it an older, darker name: Moloch. The trap is not built out of evil people. It is built out of reasonable decisions.

People have been caught in this exact trap before — and found ways out. In the 1950s, the US and the USSR tested nuclear bombs in the open air by the hundreds; neither felt able to stop. Then scientists in St. Louis collected over 300,000 baby teeth and measured strontium-90 — radioactive fallout that settles into growing bones and teeth. Levels were rising sharply in children born during the testing years, on both sides of the Iron Curtain.

In 1963, the rivals signed a treaty banning atmospheric tests. It held because such a test is almost impossible to hide: seismographs and radiation detectors meant each side could check the other. The honest footnote: testing moved underground for three more decades — the treaty banned what could be verified, not everything that was dangerous.

In 1974, biologists learned to cut and recombine DNA — and some realized their own experiments might create dangerous new organisms. They paused their own research voluntarily, before any government asked. The next year, about 140 of them met at Asilomar in California and wrote their own safety rules: which experiments could go ahead, which needed strict containment, which should wait. Genetic engineering grew up inside those rules.

Notice why it worked: the community was small, everyone knew everyone, careers depended on reputation, and pausing cost little. That recipe does not transfer easily to a race between trillion-dollar companies — AI researchers bring up this escape more often than any other, and it may be the hardest one to repeat.

In 1985, scientists found a hole in the ozone layer — the shield protecting life from ultraviolet radiation. The cause: CFCs, in fridges and spray cans worldwide. Two years later, nations signed the Montreal Protocol; it became the first treaty every single country on Earth signed on to, and the ozone layer is slowly healing.

Why did this one succeed? Replacements for CFCs already existed — chemical companies could switch products and keep selling — and satellites could measure damage and recovery for everyone to see. Stopping was cheap, and everyone could see who was keeping the promise. With carbon emissions, stopping is expensive — and the same kind of treaty has struggled for thirty years.

Every successful exit had four ingredients: a stop each side can verify; few players; a fear that is concrete and shared; and an exit that doesn't cost the quitter everything. Score today's AI race: few players — yes. Verification — genuinely hard: you cannot X-ray a company's servers, but the race runs on countable, visible data centers and chips, which some researchers call AI's potential seismograph. A shared concrete fear — not yet. A cheap exit — no: today, stopping means losing.

One more honest number: Kokotajlo puts the odds of catastrophe around seventy percent — far higher than almost anyone else. A 2023 survey of thousands of AI researchers put the median extinction risk near five percent. The experts disagree enormously about the odds. Almost none dispute the shape of the race.

The headlines track the leaderboard: whose model is smartest this month. But the leaderboard is just the trap running as designed. If the history of escapes means anything, the real signal is elsewhere: watch whether anyone is building the seismograph — the treaties, the inspection tools, the ways rivals could ever check each other's restraint. In 1963, the superpowers did not become wiser or kinder. They built a system in which stopping could be seen — and then they stopped.

Kokotajlo says he would be incredibly happy to be proven wrong, and publishes plans for how this could go well. The calm you heard at the start does not mean he has given up. It is the voice of someone who has studied the trap — and still believes in the exits.