Anthropic researcher puts odds of AI wiping out humanity above 10%

A senior researcher at Anthropic has put the odds that advanced AI eventually causes human extinction at more than 10 percent, an estimate delivered from inside one of the industry's most safety-focused labs. The researcher, who spoke on condition of anonymity to avoid a backlash, said current techniques for controlling superintelligent systems are not sufficient to rule out a worst-case outcome. The figure is not a precise forecast, they said, but a way of expressing deep uncertainty about a technology whose downside could be permanent.

The technical case behind the number rests on alignment, the problem of making a model's objectives match human values as its capabilities grow. According to the researcher, no existing method guarantees that a powerful system remains safe once it becomes more capable than the people overseeing it, and the field lacks agreed benchmarks for demonstrating that it has. That gap matters because labs are shipping larger models on shorter timelines, with Anthropic itself recently releasing Claude 4.5 and stating that it is working toward artificial general intelligence.

The warning lands awkwardly for Anthropic, which was founded by former OpenAI researchers who left over concerns about the pace of commercialization and has built its public identity around responsible development. Its proposed answer is a responsible scaling policy that ties deployment decisions to safety thresholds, but the thresholds and the enforcement behind them remain loosely defined. Estimates of existential risk inside the company have ranged widely, with some staff putting the probability far below 10 percent, so the anonymous claim reflects one researcher's view rather than a corporate position.

The researcher joins a small group of prominent figures who have made similar arguments in public. Geoffrey Hinton left Google in 2023 to speak freely about AI risk, and Eliezer Yudkowsky has called for halting large-scale training runs altogether. What is different here is the source: a working researcher at a lab whose stated mission is safe development, rather than an outside critic or a former employee.

Critics of these warnings argue they do little to change the competitive dynamics pushing capability gains, and that a vague probability offers no actionable target for engineers or regulators. Governments have begun drafting AI safety rules, but there is still no binding international agreement on testing or deployment limits, leaving labs to police themselves. The practical question is whether an internal estimate like this one shifts any deployment decision, or whether it stays a rhetorical data point.

Anthropic has not commented publicly on the specific claim, and the researcher's anonymity makes it hard to weigh against the company's own published evaluations. Watch whether Anthropic's next responsible scaling report adds concrete thresholds or third-party audits, and whether other labs follow with their own risk estimates. The broader test is whether internal numbers like this one start appearing in deployment decisions rather than only in interviews.