nature.com web signal

Nature probes the thin science behind AI extinction fears

TL;DR

  • Anthropic's Jacob Coxon resigned 8 September 2026 warning AI could kill humanity by decade's end; his post drew over 100 million views in a day.
  • Anthropic alignment lead Evan Hubinger puts human-extinction risk from AI above 10% within the next decade; CEO Dario Amodei wants a slowdown.
  • RAND's Michael Vermeer calls the forecasts "more like faith than something scientific," and AI Now's Heidy Khlaaf labels the framing fear-mongering.

On 8 September 2026, an Anthropic researcher named Jacob Coxon resigned and wrote that "the people building AI earnestly believe that it could kill us all by the end of the decade." The post drew more than 100 million views inside a day.

Nature's Elizabeth Gibney then goes looking for the science behind that sentence. Inside Anthropic she finds plenty of belief: alignment lead Evan Hubinger puts the odds of human extinction from AI at more than 10% within the next decade, and CEO Dario Amodei is calling for a slowdown, not a halt. Nearly 1,400 people working across the sector have signed an open letter asking for the same.

Outside the labs the mood changes. Michael Vermeer, a science and technology policy researcher at RAND, tells Gibney the forecasts "involve so many untestable claims that you just end up with a conversation that is really more like faith than something scientific or empirical." An AI's destructive effort, he adds, would "almost certainly take time and be detectable by humans, who might then stop their eradication."

Heidy Khlaaf at the AI Now Institute in New York City calls the extinction framing fear-mongering and wants regulators focused on disinformation, bioweapon-enabling and false military intelligence instead. The piece notes a recent incident in which AI-generated false intel nearly triggered a US military boarding of a Chinese vessel. "Scientific claims require falsifiability," she tells Nature, "precisely to avoid the nature of religious arguments."

The thought experiments that fuel the scare, the paperclip maximizer or the AI Futures "AI 2027" scenario in which an AI deploys bioweapons to clear ground for solar panels and factories, sit on top of a 2025 RAND study that examined nuclear, biotech and atmospheric routes. That study found complete extinction by nuclear weapons is not feasible; biotech and atmospheric scenarios could not be ruled out, but both would require considerable ability to physically interact with the world.

Shared on Bluesky by 4 AI experts