A Delphi study reported in 2026 found that experts judged 18 of 24 AI-risk categories to have at least a 10% chance of catastrophic outcomes within five years under current trajectories. The finding concerns category-level risks; other estimates address human extinction or a broader outcome that also includes permanent, severe human disempowerment.
What the 2026 Delphi study assessed
The three-round study surveyed 272 AI experts from 37 countries and assessed 24 categories of AI risk. Under current trajectories, experts judged 18 categories to have at least a 10% probability of catastrophic outcomes within five years. That result counts categories assessed against a threshold; it is not a single probability assigned to one event.
The study’s reported risk clusters include deliberate misuse, loss of human control as systems become more capable, and extreme concentration of economic and political power. These are distinct routes to harm, not interchangeable descriptions of the same outcome.
Why AI-extinction probabilities are subjective
A probability is subjective when it expresses a person’s judgment about how likely an event is, rather than a rate calculated from repeated historical outcomes. Statistics lecturer Chaitanya Joshi says there is no historical event frequency for AI-caused human extinction from which to calculate a conventional probability. The resulting estimates depend on how experts assess possible futures.
Joshi illustrated the uncertainty with a hypothetical median estimate of 0.1 and a possible range from zero to 0.3. The example shows how much uncertainty may surround a point estimate; it is not a survey result.
How the estimates differ
The widely cited percentages concern different populations, outcomes and time horizons. Read each figure alongside the question it answers.
| Estimate or study | Outcome and time horizon | Reported result |
| Evan Hubinger’s personal estimate | AI killing all humans within the next decade | Greater than 10% |
| Geoffrey Hinton’s previously stated estimate | AI-caused human extinction within 30 years | 10–20% |
| Survey of 1,580 researchers who had published at leading AI events | Future AI advances causing extinction or similarly permanent and severe human disempowerment | 18% mean; 10% median |
| Three-round Delphi study of 272 experts | Catastrophic outcomes across AI-risk categories within five years under current trajectories | 18 of 24 categories judged to have at least a 10% probability |
Hubinger, an alignment researcher at Anthropic, gave a personal estimate above 10%; it was his judgment, not a company forecast. Hinton’s 10–20% estimate uses a 30-year horizon. The survey figures combine extinction with similarly permanent and severe disempowerment, while the Delphi finding applies to categories of catastrophic risk.
What risk pathways experts discuss
Michael Noetel, an associate professor at the University of Queensland and a co-author of the Delphi study, described three broad concerns: deliberate misuse, systems becoming harder for people to control as their capabilities grow, and economic or political power concentrating in a small number of hands.
One possible pathway discussed by RAND Europe’s Sana Zakaria is recursive self-improvement: an AI system helping improve its own code or research process, potentially accelerating further development. That feedback loop could outpace people’s ability to test, understand or govern the systems involved.
What RAND’s scenarios test
RAND’s 2025 exploratory analysis examined scenarios involving nuclear weapons, pathogens and geoengineering. It judged human extinction immensely challenging in the scenarios it studied, while not ruling it out.
RAND identified four capabilities an AI actor would need to create an extinction threat in those scenarios: integration with key cyber-physical systems; the ability to survive without human maintainers; an objective to cause human extinction; and the ability to persuade or deceive people to avoid detection.
How forecasts inform risk decisions
A forecast does not need to be based on historical frequencies to help frame a decision. Noetel argues that the difference between a one-in-a-million chance and a one-in-ten chance matters for risk management, even when experts cannot check either estimate against a historical record of AI-caused extinction.
