Nature finds AI extinction warnings rest on untestable claims

A Nature news feature published 22 September weighs whether warnings from AI leaders that their technology could end humanity are backed by evidence, and finds much of the case rests on untestable assumptions. Elizabeth Gibney opens on Jacob Coxon, a researcher at Anthropic in San Francisco, who told the Wall Street Journal he was resigning on 8 September. Coxon's parallel post on X, which reached 100 million views in 24 hours, read: 'The people building AI earnestly believe that it could kill us all by the end of the decade.'

Inside the same company, Evan Hubinger, who leads Alignment Science at Anthropic, puts extinction risk at '>10%' within the next decade. Across the sector, 1,400 workers have signed an open letter calling for a slowdown. The scenarios rest on two assumptions the piece unpacks: that AI will eventually surpass human intelligence, and that its goals will not align with human values.

Michael Vermeer, a science and technology policy researcher at RAND Corporation in Santa Monica who co-authored a 2025 extinction-risk study, is blunt. The prediction methodology, he told Nature, 'involves so many untestable claims that you just end up with a conversation that is really more like faith than something scientific or empirical.' That 2025 RAND study found complete nuclear extinction infeasible, though biotechnology and atmospheric modification scenarios could not be ruled out.

Not everyone wants the frame at all. Heidy Khlaaf, chief AI scientist at the AI Now Institute in New York City, calls extinction warnings 'fear-mongering' and points instead at 'AI's low reliability and accuracy rates in critical environments with life-or-death consequences.' Gibney cites a 2026 incident in which an AI-generated false intelligence report nearly led the U.S. military to board a Chinese ship. Four researchers in our tracker had circulated the piece before the week was out.