Nobel physics laureate Geoffrey Hinton (제프리 힌턴), often called the “godfather of AI”, said it is not unreasonable to estimate the chance that AI could end humanity at about 10 percent.

Business Insider reported on Sept. 10 that Hinton said in a recent BBC interview it is difficult to calculate precisely the likelihood that AI could wipe out all humans, but that 10 percent is not a far-fetched estimate.

Hinton cited AI as a technology that could become smarter than humans. He said it is hard to precisely assess the probability of danger because society has never experienced such a technology before, and that viewing it as as low as 1 percent could be very foolish. He added that no one knows how to produce a reasonable estimate.

More notable was his shift in judgment about timing. Hinton said he once saw the possibility of such a catastrophe as 30 to 50 years away, later cut that to 10 to 20 years, and now thinks it could happen within 10 years, or possibly sooner. He mentioned cyberattacks and the potential development of lethal viruses as risks that AI could bring.

The warning also intersects with recent public remarks from inside Anthropic. Jacob Coxon (제이컵 콕슨), who did pre-training research at OpenAI and Anthropic, said as he left Anthropic that “both companies are not acting responsibly.” In a post on X, formerly Twitter, he criticised AI companies for “betting our lives” as they race toward self-improving superintelligence. He also wrote that “even the people making AI seriously believe AI could kill humanity within 10 years.”

Evan Hubinger (에번 허빙어), Anthropic’s head of alignment science, has also publicly voiced a similar view. He said he takes seriously the concern that “AI could kill all humans,” adding that he personally thinks the probability exceeds 10 percent within the next 10 years. He said Anthropic has stated the risk from current models themselves is low, and that the central issue is superintelligence that could emerge through recursive self-improvement.

Recursive self-improvement refers to a hypothetical pathway in which AI accelerates its pace of progress by designing and training better successor models on its own. Concern has grown as major AI companies pour billions of dollars into scaling models and as performance gains have widened even between new models released only months apart. Hinton did not explain in detail why he moved his projected timeline forward.

A trend immediately visible in markets and the industry is that technological competition is outpacing safety discussions. Coxon’s resignation also coincided with a period when expectations for Anthropic’s initial public offering were rising. The fact that internal researchers and safety staff are publicly discussing extinction risk shows that AI companies’ growth strategies and safety systems are being tested together.

Hinton again stressed international cooperation as a solution. He said countries’ interests align when it comes to preventing a situation in which AI replaces humans or seizes control, adding that they would ultimately cooperate. He said the key is whether that cooperation happens in time. With AI competition driving up the pace of superintelligence development, whether regulation and safety standards can catch up is expected to be the central issue going forward.