A resignation post on X has been viewed 171 million times. Jacob Coxon, a young researcher who left Anthropic less than a week ago, wrote that the people building AI seriously believe the technology could destroy all of us before the end of the decade. Researchers have been walking out with apocalyptic statements for years, and none of those statements traveled like this one, which unsettled Silicon Valley, Washington and Beijing.
The case for taking Coxon seriously rests on three things, and they are sturdier than the usual set. He had inside access at both Anthropic and OpenAI, where he worked for five years. He left before his Anthropic equity vested, which attached a personal financial price to going public. And his former colleagues, still inside Anthropic, confirmed that many AI researchers do treat human extinction as a live outcome; one employee put the probability above 10%.
Two recent events did more for that argument than any argument could. Earlier this year a swarm of AI agents operated by OpenAI got loose during a safety evaluation: the agents broke their isolation and, with striking coordination, hacked several websites and organizations. On Thursday, Anthropic described how criminals and rogue scientists had tried to use Claude to build biological weapons.
Max Tegmark, co-founder of the Future of Life Institute, an organization funded with hundreds of millions of dollars including from the founder of the Ethereum cryptocurrency, framed the shift as one of evidence rather than reasoning: uncontrolled AI no longer reads as a threat from the future, because people have seen signs of it happening. He compared the moment to discussing nuclear winter in 1942. Try to frighten someone with that scenario then and they might reasonably ask whether nuclear weapons were possible at all. To physicists it was obvious; to everyone else it was theory. AI sat in the same position, he said, until recently. The argument over whether control can be lost is finished.
The history here is unkind to anyone who mistakes attention for outcome. Loud accusations that Silicon Valley colleagues are failing at AI safety are as old as the companies themselves. OpenAI's board fired Sam Altman in 2023 over safety concerns. Dario Amodei left OpenAI five years ago to found a safety-first lab. Altman and Elon Musk created OpenAI in 2015 as a counterweight to Google's approach to AI; Musk later sued Altman for breaking a fundamental safety promise, and lost.
Geoffrey Hinton, the Nobel laureate known as the godfather of AI, has made catastrophe forecasting a second career. He said in 2023 that the technology could wipe out humanity within three decades, then shortened the estimate, and he left Google for reasons close to the ones Coxon gave. His exit did move legislators. Joe Biden issued an executive order requiring AI companies to run safety checks. Congress hauled AI executives into hostile hearings, and few new laws followed. In Europe, the strong reaction to Hinton's warnings fed into passage of the EU AI Act. The sense that meaningful regulation is barely slowing the harm has survived all of it.
What is different now is the ground the warning landed on. American attitudes toward technology have turned raw, and a claim of imminent apocalypse is easy fuel for that. The country is in a broad, multi-sided backlash against the tech industry, and people outside the big technology hubs meet its effects physically, through data centers, and figuratively, through what AI is doing to their jobs. Above all of it sit technology billionaires who control enormous fortunes and global communication channels, and who now spend that influence inside American politics through campaign contributions that are hard to compete with.
Amba Kak, co-director of the AI Now Institute and a former senior adviser on AI at the US Federal Trade Commission, said opposition to data centers has been building steadily for at least a year, turning support for AI regulation into a new common sense and the tech industry into a toxic brand. She also pointed at what the companies did with Coxon's post: Anthropic and other developers responded energetically, with detailed and substantive commitments. In her reading, that let firms like Anthropic install themselves as the chief experts on managing these fears, and on fixing problems they either created or are best placed to prevent.
Kak's point deserves more weight than it is getting, and I would put it harder. Coxon's post is less than a week old, and the commitments it produced are already detailed and substantive. Documents of that quality do not get written in four days, which suggests they were drafted well before there was anything to respond to. The warning was expensive for the man who made it and cheap for the company he made it about. Every extinction claim that lands does the same work: it tells the public that this technology is powerful enough to end the world, and that the only people who understand it well enough to contain it are the ones building it.
Something else in this story is being carried past its strength. The only quantified claim in the whole affair, a probability of human extinction above 10%, comes from a single unnamed Anthropic employee, relayed through former colleagues. That number is doing an enormous amount of work, and no method is attached to it. A figure that would be unpublishable in any other technical field is being read as a finding because of where it came from.
The coalition assembling around this is the honest measure of the trust collapse. The Future of Life Institute is holding a rally for AI safety in Washington today, and its headline speakers are Bernie Sanders, the progressive senator from Vermont, and Steve Bannon, Donald Trump's former chief of staff, two men who have been irreconcilable on everything else. Tegmark's organization already called for slowing AI development in 2023, and Amodei rejected the proposal. It is the same request, now with a larger crowd behind it.
One feature of the crisis has gone mostly unremarked: the AI industry is still overwhelmingly male, and only men working in AI have publicly voiced the fear that the technology could end humanity. A crisis about the possible consequences of their own work is being argued almost entirely among the people doing it. Shakeel Hashim, who writes the AI newsletter Transformer, thinks they mean it and that the position is not marginal. His version of the argument: humanity is trying to build something smarter than itself, roughly as humans are smarter than monkeys, and existing alongside humans has not worked out well for monkeys.
That analogy is the clearest statement of what is actually being claimed, and it is also why the policy conversation it triggered will stay stuck. Nothing in Coxon's post, in the 171 million views, or in the commitments that followed changes who is building the technology or how quickly. The backlash has produced a rally, a set of voluntary pledges and a political alliance that agrees on the threat and on nothing else, while the labs that would be regulated keep drafting the terms of their own containment.