Nature probes thin science behind AI extinction warnings

A Nature news feature by Elizabeth Gibney tests the empirical basis of warnings from senior AI-lab figures that the technology could wipe out humanity, and finds much of the case rests on scenarios independent researchers call untestable. The piece was prompted by the September 8 resignation of Anthropic researcher Jacob Coxon, who wrote that "systems the company is developing could spiral out of control and destroy humanity." Evan Hubinger, who leads alignment science at Anthropic, reposted the message and attached his own figure: a risk of human extinction ">10% within the next decade."

Chief executive Dario Amodei, Nature reports, has called for the field to slow rather than halt frontier development, and nearly 1,400 AI-sector workers have signed an open letter backing a slowdown. Four researchers we track reposted the piece after it went up.

Independent scientists quoted by Gibney are unconvinced. The forecasts, Michael Vermeer of the RAND Corporation tells Nature, involve so many untestable claims that the conversation becomes "more like faith than something scientific or empirical." A 2025 RAND study he worked on judged complete extinction via nuclear weapons not feasible; biotechnology and atmospheric modification "could not be ruled out," but each would need physical capacities current systems do not have, and would likely be detectable in advance.

Heidy Khlaaf, chief AI scientist at the AI Now Institute, dismisses the extinction framing as fear-mongering and argues regulators should focus on the technology's "low reliability and accuracy rates in critical environments," alongside disinformation, bioweapon facilitation and false military intelligence.