Evan Hubinger has spent years working on one of the hardest problems in artificial intelligence: how to make sure increasingly capable AI systems actually do what humans want, rather than pursuing goals that diverge from human interests in ways nobody intended. As the Alignment Science Lead at Anthropic, one of the world’s leading AI labs, his job sits at the center of the industry’s most consequential open question. This week, he put a number on the risk that keeps him up at night, and that number startled a lot of people.
A Blunt Estimate, Publicly Stated
Hubinger’s comments came in response to a former Anthropic colleague, Jacob Coxon, who announced his resignation from the company and accused both Anthropic and OpenAI, where he had previously worked, of acting irresponsibly. Coxon argued that neither company was behaving responsibly, describing them as racing toward self-improving superintelligence while gambling with humanity’s future.
Rather than distancing himself from that framing, Hubinger largely agreed with it. Posting on X, he said he genuinely believes AI could kill all of humanity, and that he personally estimates that risk at above 10% within the next ten years. He added a candid admission about his own employer’s limitations: Anthropic is trying its best, he said, but the company does not yet have a plan to solve alignment for superintelligent systems and is not clearly on track to develop one.
It is a striking thing for a senior researcher to say about the company that employs him, and it quickly drew attention well beyond AI safety circles, landing coverage from outlets ranging from CBS News to RT.
Not About Today’s Chatbots
Importantly, Hubinger’s warning is not a claim that the AI models currently available to consumers pose an imminent threat. According to reporting on his remarks, he indicated the immediate danger posed by present-day AI systems is comparatively low, with his real concern centered on where the technology is headed rather than where it currently stands. The specific fear driving his estimate is recursive self-improvement: the prospect of an AI system capable of designing and refining its own successors without meaningful human oversight. That capability does not exist yet, but researchers at both Anthropic and OpenAI have separately warned that its arrival could make it substantially harder for humans to retain control over advanced systems.
It’s also worth noting that Hubinger was careful to frame his figure as personal rather than institutional. Coverage of his statement has emphasized that the 10% estimate reflects his own judgment, not an official forecast issued by Anthropic as a company. That distinction matters: it says less about Anthropic’s formal risk modeling and more about the private calculus of the people building the technology.
Part of a Broader Pattern of Alarm
Hubinger’s post did not emerge in isolation. He was one of roughly 1,400 AI researchers who signed the “Pacing the Frontier” open letter over the summer, which called on the U.S. government to help develop tools capable of deliberately slowing the pace of automated AI development. More than a thousand employees across AI companies had already put their names to similar warnings earlier in the year, arguing for international cooperation on governance mechanisms.
Coxon’s resignation and Hubinger’s response have also landed at a moment when Washington is actively debating how to regulate the industry. A bipartisan proposal in the House, sometimes referred to as a “kill switch” bill, would give Congress authority to shut down AI models judged to threaten public safety, while separate legislative efforts have aimed to establish broader frameworks for governing advanced AI deployment.
Why It Resonates
What makes Hubinger’s statement notable isn’t that a critic of the AI industry raised an alarm — those warnings are common. It’s that the alarm came from inside one of the labs most responsible for pushing the technology forward, from someone whose entire job is to reduce exactly the risk he’s describing. When the person leading alignment research at a frontier lab says his own company lacks a clear plan to keep future superintelligent systems under control, it reframes the AI safety debate from an abstract, hypothetical concern into something the people closest to the technology are actively wrestling with, in public, in real time.
