An AI researcher said AI could kill us all. What do the numbers say?

The warning went viral. A 2026 panel of 194 experts put a sobering number on one worst-case scenario—but it is conditional, disputed and not a prediction that humanity will end.

On September 9, Jacob Coxon said he was resigning from Anthropic after working in AI research at Anthropic and OpenAI. He argued that the companies were moving too quickly and warned that advanced AI could threaten human life. The Associated Press reported that his posts reached more than 100 million people overnight and that two current Anthropic employees agreed with his concerns. That makes it a widely shared warning from people close to the work. It does not make it a settled scientific forecast or an official statement from either company.

There are forecasts to examine, but the question and the definition matter. A Longitudinal Expert AI Panel report released in June 2026 collected forecasts from 194 experts. It defined an “AI-related global catastrophe” as an AI-caused event in which more than 10% of the world’s population dies within five years. The median expert estimated a 10% chance of that threshold by 2100 if AI capabilities progress rapidly, compared with 2% under a slow-progress scenario.

That 10% is not “a 10% chance AI kills everyone.” The survey’s threshold is more than one in ten people, not human extinction; the estimate depends on a scenario that has not happened; and a panel’s median judgment is a forecast, not a measured frequency. Experts also disagree about how likely the fast-progress scenario is, what counts as AI-caused and whether AI may reduce other risks.

The same report asked a different, lower-threshold question: whether AI would cause at least 50 deaths or $100 billion in damage by 2050. Experts gave that event a 62% median probability. It can sound like a contradiction until you notice the event is defined differently: a serious incident is not the same as a global catastrophe, and neither is the same as extinction. Headlines often compress all three into “AI could end the world.”

So what should you take from the viral warning? Not “nothing to see here,” and not “the end is scheduled.” Treat the resignation as a reason to inspect the evidence, then check every risk number for its population, time horizon, scenario and definition. The honest answer is that experts see serious possibilities and remain deeply uncertain about how likely they are.

Before sharing an AI-doom statistic: ask what event it counts, by when, under which scenario—and whether it is observed data, an expert forecast or one person’s warning.

Forecasts describe respondents’ judgments under specified questions and scenarios; they are not established probabilities or guarantees.

Sources: Associated Press, Anthropic researcher resigns with warning about AI development; Forecasting Research Institute, Longitudinal Expert AI Panel Wave 9; International Delphi study of 272 AI experts, 2026.