An AI Researcher Just Quit Anthropic Over Existential Fear. Should the Rest of Us Be Worried?

Key points

• Jacob Coxon, an AI safety researcher, publicly resigned from Anthropic, one of the leading AI companies in the US, in a post on X in 2025.
• Coxon claims colleagues are not staying because they doubt the danger, but because they fear rivals with fewer safety concerns will win the race instead.
• The concern centres on "self-improving superintelligence", AI that could redesign and upgrade itself without human involvement.
• Insiders reportedly talk openly about a "crunch time", a window of one to two years they believe will determine how this technology ends up.
• Anthropic has not publicly responded to Coxon's claims.

AI safety researcher Jacob Coxon recently resigned from Anthropic, the San Francisco lab behind the Claude family of AI chatbots, posting a public statement on X saying he cannot keep working for companies "gambling with our lives". It landed differently from the usual warnings. Louder names have said scarier things. This one stuck.

What exactly is he worried about?

The specific fear is self-improving superintelligence: AI that is not just very clever, but capable of designing a smarter version of itself, then a smarter version of that, looping forward without any human in the chain. Nobody has built this yet. The worry is about what happens if someone does.

Coxon's most unsettling claim is not that his former colleagues think he is wrong. It is that many quietly agree with him, and keep working anyway. The reason, he says, is a kind of grim logic: if powerful AI is coming regardless, better that cautious people build it than people who do not care about the risks at all.

That is a genuinely hard argument to dismiss.

Inside the lab, staff apparently talk regularly about "crunch time" or "the endgame", shorthand for the belief that decisions made in the next year or two will shape how this technology lands for everyone. Whether that is accurate or a form of self-importance common in Silicon Valley is genuinely unclear. Both things can be true at once.

So should ordinary people panic?

Not panic, no. But paying attention is reasonable.

The Guardian columnist Gaby Hinsliff, whose piece first drew wider attention to Coxon's resignation, raised a fair point: it is difficult to know whether warnings like this are genuine alarm bells or a strange kind of marketing, an industry making its products sound terrifyingly powerful to attract investment. AI companies have done that before.

What is harder to wave away this time is the specific detail. Coxon is not an outsider speculating. He worked there. And his description of an industry "begging to be saved from itself" matches a pattern: the same labs racing to build the most powerful AI are also the ones publishing safety research and calling for regulation.

For most people, the practical takeaway is not to change daily life, but to treat AI policy like any other issue that affects you, which means expecting politicians to take it seriously and asking them why they are not when they don't.

Common questions

What is Anthropic and why does it matter?

Anthropic is one of the best-funded AI labs in the world, valued at around $61 billion as of early 2025. It built the Claude chatbots and positions itself as focused on AI safety, which makes Coxon's resignation from inside that specific company notable.

Is this the same as previous AI doom warnings?

Broadly similar fears have come from bigger names, including Geoffrey Hinton and Sam Altman. What makes Coxon's account different is the claim that colleagues share his fears but feel trapped by competitive pressure, rather than disagreeing with his assessment.

Can anything actually stop this research?

Some researchers and politicians call for binding international agreements, similar to nuclear non-proliferation treaties. No such agreement exists yet for AI. That gap is exactly what concerns people like Coxon.