Nature tests AI extinction claims from Anthropic staff

Anthropic researcher Jacob Coxon resigned on September 8 saying he feared the systems his company builds could destroy humanity, and Nature has now published a science-desk piece testing that claim against the underlying evidence. Coxon's post on X, that "the people building AI earnestly believe that it could kill us all by the end of the decade," drew more than 100 million views in a day. Hours later Evan Hubinger, who leads Anthropic's alignment science, reposted it and added his own extinction estimate of ">10% within the next decade." Nature's Elizabeth Gibney walks through who is making these claims and what the science actually shows.

The scenarios industry insiders point to are speculative. A forecast called AI 2027, from the non-profit AI Futures project, imagines an AI unleashing a biological weapon to clear humans out of the way of solar panels and robot factories. Michael Vermeer, a science and technology policy researcher at the RAND Corporation, is unimpressed with the genre. Making such predictions, he tells 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."

RAND's own 2025 study looked at the three concrete technologies through which an AI might in principle wipe out the species: nuclear weapons, biotechnology, and deliberate modifications to the atmosphere. It found complete extinction by nuclear weapons is not feasible; the biotech and atmospheric scenarios "could not be ruled out." That is a narrower and more sober picture than the >10% figure being posted on social media.

Heidy Khlaaf, chief AI scientist at the AI Now Institute, is harder on the extinction framing itself, which she calls "fear-mongering." Her concern is nearer-term: "AI's low reliability and accuracy rates in critical environments with life-or-death consequences." Four of the researchers we follow in our directory circulated the piece the day it ran, and it lands as a fight over which risks deserve regulatory oxygen: the speculative extinction ones being amplified from inside frontier labs, or the operational failures already showing up in deployed systems.