Scientists call AI extinction warnings 'more like faith'

On September 8, Anthropic researcher Jacob Coxon resigned and wrote that 'the people building AI earnestly believe that it could kill us all by the end of the decade.' Anthropic's alignment science lead, Evan Hubinger, amplified the post and put the number at '>10% within the next decade.' The repost drew over 100 million views in 24 hours, Nature reports in a piece by Elizabeth Gibney.

The problem, independent researchers tell Nature: the warnings don't have much science underneath them. Michael Vermeer, a science and technology policy researcher at RAND Corporation, told Gibney that 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 of extinction scenarios found complete extinction by nuclear weapons was 'not feasible,' though biotechnology and atmospheric modifications 'could not be ruled out.' Vermeer also notes that an AI's destructive efforts would 'almost certainly take time and be detectable by humans, who might then stop their eradication.'

Heidy Khlaaf, chief AI scientist at the AI Now Institute, is blunter. She calls the existential framing 'fear-mongering' and points to what current models actually are: 'largely probabilistic systems' lacking human-like understanding. Her concern is the harms the field is already documenting. CNN reported this year that false information in an AI-generated report almost led the US military to board a Chinese ship.

The timing is loaded. Anthropic CEO Dario Amodei has called for a slowdown rather than a halt, citing recent cybersecurity incidents and recursive self-improvement. Sam Altman and Elon Musk have joined the slowdown chorus. Senator Bernie Sanders has proposed legislation banning artificial superintelligence. Around 1,400 AI-sector employees signed an open letter. David Sacks, co-chair of the US President's Council of Advisors on Science and Technology, said firms face 'massive product-liability exposure' if models enable damaging cyberattacks. Anthropic itself is reportedly approaching an IPO.

Four of the researchers we track in our Who's Who directory flagged the piece, which hints at where this debate is actually landing: in the policy fight over which category of risk, speculative or documented, regulators should price in.