ChatGPT Whistleblower: They Offered Me $2m To Keep Quiet, No One Is Ready For What's Coming!

Two hours on the podcast that gave Roman Yampolskiy his 99% headline, and the title gets the money story backwards. Kokotajlo did refuse to sign OpenAI’s non-disparagement clause on his way out in 2024, and refusing meant forfeiting equity worth roughly 80% of his family’s net worth. But he kept it: the refusal became a public scandal, employees started asking in Slack why their equity could be clawed back, and the company reversed the paperwork. Leadership said they hadn’t known. Asked whether he believes that, he says no.

The second correction is to his own famous number. He has been quoted at a 70% chance of human extinction; what he actually holds is 70% that it goes horribly wrong — AI takeover being one branch of that, and takeover not necessarily meaning everyone dies. The distinction matters to him enough that he makes it unprompted.

What is new here is AI 2040: Plan A, the follow-up to AI 2027 and the reason to watch this one if you have already watched the others. AI 2027 was a prediction; Plan A is a recommendation, and he is blunt that the two are not the same document — “we don’t expect the world to listen to us.” In the scenario, regulation arrives in 2029, development continues at a slower and more transparent pace spread across more countries and companies, and superintelligence arrives in 2040 rather than around 2030. The labelling is deliberate: plans B, C and D are the alternatives, with D being AI 2027’s race continuing unchanged. The graph he walks through makes the timing argument sharper than the prose does — wait until most people have lost their jobs before regulating, and the companies already have superintelligence, because getting there first is their strategy.

His account of the labs is not that anyone is lying. It is rationalization: each CEO has convinced himself that things will probably be fine and that the worse outcome is someone else getting there first. He grants Anthropic has been more willing to say and do things that cost it money and standing, then refuses the frame that invites — “I really don’t want to be in a situation where we’re like, which CEO is the least bad CEO.” Nobody should be trusted with that much power.

The interview’s hardest stretch is about children. He and his wife had their first in 2019, before his timelines shortened; afterwards he told her he did not want more, and then gave in. He is still not sure. The reason to have kids, he says, is largely about the future.

These are auto-captions, so nothing here is quoted verbatim.