Nature dissects AI extinction warnings after Anthropic exit

Anthropic's alignment science lead, Evan Hubinger, puts the chance of AI killing all humans within the next decade at greater than 10 percent.

Nature's news team sets that figure next to the resignation of his colleague Jacob Coxon, who quit Anthropic on September 8 and posted on X that the systems he had helped build could "spiral out of control and destroy humanity." The post drew more than 100 million views in 24 hours. Nearly 1,400 people in the AI sector have since signed an open letter calling for a slowdown; Anthropic chief executive Dario Amodei is among those arguing for slowing frontier development rather than halting it.

"The people building AI earnestly believe that it could kill us all by the end of the decade," Coxon wrote.

The question Nature's science desk puts to that belief is whether it rests on anything testable. Michael Vermeer, a researcher at the RAND Corporation in Santa Monica, said extinction 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." A 2025 RAND study walked through the standard scenarios: AI-triggered nuclear war was judged "not feasible" as an extinction mechanism, while engineered pathogens and atmospheric modification "could not be ruled out," though any such path would "almost certainly take time and be detectable by humans."

Heidy Khlaaf, chief AI scientist at the AI Now Institute in New York, calls the extinction framing "fear-mongering" and argues the attention belongs on AI's low reliability and accuracy in critical deployments today. The paperclip maximiser, and the widely-circulated AI 2027 forecast in which a misaligned model "unleashes a biological weapon to kill humans" to clear ground for solar panels, remain thought experiments.