Four resignations that should change how we talk about AI !!
On the morning of September 9, 2026, a young researcher walked away from one of the most coveted jobs on the planet. Nobody fired him. Nobody poached him. He posted a short thread on X, thanked his colleagues, and announced he was leaving not just his company but his entire industry, at the exact moment that industry is minting more wealth than any in history.
His reason: he believes the thing he was helping build could kill everyone.
Easy to dismiss as one burnt-out employee, until you see what happened next. Within hours, a senior leader at the same company responded in public. Not to deny it. To confirm it. The man whose actual job title is Alignment Science Lead, the AI safety executive responsible for keeping this technology under control, wrote that he and his colleagues "earnestly believe AI could kill all humans" and put his personal odds above 1 in 10 within the next decade. He added that his company does not yet have a plan to solve the problem and is not clearly on track to find one.
This was not a weapons manufacturer. Not a virology lab. The researcher was Jacob Coxon, the leader was Evan Hubinger, and the company was Anthropic, maker of the Claude chatbot, a firm that positions itself as the safety-focused AI lab. The one that was supposed to be careful.
And here is the part that should hold your attention: this was not the first walkout. It was the fourth in a pattern stretching back three years.
A pattern, not an incident
May 2023. Geoffrey Hinton, the Turing Award winner the field calls the Godfather of AI, resigns from Google specifically so he can warn the public without a corporate filter. He initially estimates a 10% chance that AI causes human extinction. By December 2024 he raises it to 10 to 20% within 30 years, comparing our future position to a three-year-old trying to control an adult.
April 2024. Daniel Kokotajlo, a researcher at OpenAI, resigns over safety concerns and is handed one piece of paper. Sign this non-disparagement clause, keep your vested equity. Refuse, and forfeit roughly $2 million, about 85% of his family's net worth by his own accounting. He refuses. He decides his right to warn the public is worth more than nearly everything his family owns. (OpenAI later reversed the policy under public pressure and returned the equity. Remember that detail. It matters at the end of this article.)
February 2026. Mrinank Sharma, head of Anthropic's Safeguards Research team, an Oxford PhD leading the very group built to contain these risks, resigns with a public letter declaring that "the world is in peril" and that he repeatedly watched commercial pressure push aside what matters most. He leaves to write poetry. The letter crosses a million views in a day.
September 2026. Coxon walks, telling researchers still inside that the labs are "gambling with our lives", and that executives soften their language for the press while expressing fear privately.
Four exits. Three companies. One consistent message, delivered by the people with the deepest inside view, several of whom paid real money or real careers for the right to say it.
What the numbers actually say
You do not have to take any single person's estimate on faith. In the largest survey of AI researchers ever conducted, 2,778 authors published at the field's top venues, the median researcher put a 5% probability on advanced AI causing human extinction or permanent human disempowerment. Thirty-eight percent put it at 10% or higher.
Five percent is the optimistic middle of the field. Nobody boards a flight with a 5% chance of crashing. Hinton says 10 to 20. Hubinger says above 10 within a decade. Kokotajlo, who forfeited a fortune for the right to speak, puts the odds of catastrophe near 70%, a category he defines to include outcomes he considers as bad as extinction, such as a small group using super-intelligent AI to lock in permanent, irreversible power over everyone else.
The time frames and definitions differ across these numbers, and that difference matters more than any single figure. Nobody is claiming a shared, precise forecast. What they share is direction: every named estimate sits meaningfully above the field's own median.
One honest caveat, straight from Hubinger himself: today's chatbots are not the danger. The risk he flags is tomorrow's systems, AI that improves itself faster than humans can supervise. This is not fear of what exists. It is fear of what is being raced toward, openly, on public timelines of one to four years.
The trap: why nobody stops alone
Here is the question everyone asks: if the builders believe this, why do they keep building?
Coxon answered it on his way out, and his answer is the most important paragraph in this story. At Anthropic, he said, the risks are well understood internally. But the company has convinced itself it has no choice but to keep racing, because if it slows down, a less careful competitor reaches superintelligence first.
Every lab reasons the same way. Each one fears the others crossing the line sooner, so each one runs faster, which makes every other lab run faster still. No single company can stop alone without handing the future to whoever kept going. These are not villains. They are intelligent people locked in a race where the individually rational move is collectively suicidal. Which leads to the only conclusion the logic allows: they all have to slow down together, or none of them will.
Together does not happen from inside. It has to be arranged from outside.
What the insiders are actually asking for
Not one of these people asked us to panic, and none of them called for burning the industry down. Their asks are specific. Coordination among the labs. Legal protection for employees who warn the public, the demand of the Right to Warn letter signed by 13 OpenAI and DeepMind insiders in June 2024. A temporary halt on capability jumps if coordination fails, per Coxon. A negotiated international slowdown, per Kokotajlo's Plan A. Reasonable, bounded, achievable things. What they lack is not a plan. It is public pressure behind the plan.
The lever already worked once
Recall the Kokotajlo detail. When he refused to sign, one journalist investigated, employees revolted, and within about a week OpenAI reversed a years-old policy, returned his equity, and freed every former employee from those gag clauses. That reversal is the reason Sharma and Coxon could speak openly in 2026 at all. Public attention did not just win once. It created the conditions for every warning that followed.
That is the lever, and it is in your hands, not theirs. When 2,407 US adults were polled, 58% already agreed that AI extinction risk should be a global priority alongside pandemics and nuclear war. Yet 68% said they barely worry about AI day to day.
The public is not divided on this. It is quiet.
The people closest to this technology gave up fortunes and careers to start this conversation. The least the rest of us can do is have it. Talk about it at dinner. Share what the insiders actually said. Ask the leaders you follow where they stand. Companies bend when enough people watch. We have already seen it happen.
Comments