Survey: AI Researchers Put 18% Odds on Human Extinction in 2024

More than 1,500 leading AI researchers estimated an 18 percent average probability that AI leads to human extinction, according to survey data from 2024. The finding surfaced in a debate kicked off by Anthropic researcher Jacob Coxon, who argued on X that the risk deserves more serious attention. The number is not a fringe estimate from outsiders: it comes from people building the systems at the center of the industry.

The online argument drew in prominent voices. OpenAI researcher Daniel Selsam described the situation as a ticking time bomb, while a former DeepMind researcher said AI could kill everyone. Their willingness to attach their names to those warnings marks a shift from the more cautious language typically used in official company statements and research papers.

The 18 percent figure is an average across respondents, not a consensus prediction. In survey work of this kind, responses usually range from near-zero to well above 50 percent, so the mean can mask deep disagreement. What stands out is that the midpoint sits far above what most risk models treat as acceptable for a technology already deployed in products used by hundreds of millions of people.

AI labs have signed on to safety frameworks in recent years and published frontier safety policies with capability thresholds. Those documents generally focus on misuse and loss of control, but they do not assign explicit probabilities to human extinction. A survey average of 18 percent, taken at face value, would place the technology in a risk category regulators reserve for hazards with far more established evidence.

The figure from 2024 is not static. Reporting on the survey notes that the estimated probability has continued to climb since the data was collected. That trend matters because it reflects changing views among the same technical community whose judgments shape model evaluations, safety testing and, increasingly, government rulemaking.

What happens next depends on whether these views translate into action. Watch for follow-up surveys that track the number over time, for whether major labs publish their own internal risk estimates, and for whether the gap between researchers' private assessments and public corporate messaging narrows or widens as new models ship.