P doom explained: who has said what about AI extinction odds, and how to read the number

P doom is the probability a person assigns to AI causing human extinction or an equally catastrophic outcome. It is a gut estimate, not a measurement, and the published numbers run from effectively zero to above 99%. This page explains what the figure means, lists the public estimates with their dates as of October 1, 2026, and says how much weight each deserves.

What p doom means

The term is short for "probability of doom." There is no standard definition. Some people mean extinction, some mean a permanent loss of human control, some mean any global catastrophe, and some add a time limit such as "by 2100." When two people trade numbers, they are often answering different questions.

The number is a subjective judgment. Geoffrey Hinton, reported in a September 2026 BBC Newsnight interview, framed his 10% as a personal guess made without a mathematical framework, in a situation with no historical precedent. That framing applies to nearly every figure below. For the underlying question, see our AI takeover explainer.

The one number from a survey

The only large-sample figure comes from researchers, not celebrities. The 2023 AI Impacts survey of 2,778 AI authors, published on arXiv on January 5, 2024, found that "between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction." Even among net optimists (68.3% of respondents), 48% gave at least a 5% chance of extremely bad outcomes. The same survey put a 50% chance on machines outperforming humans at every task by 2047, a date we track alongside others in the AGI timeline.

Public estimates, dated

These come from Wikipedia's p(doom) compilation, opened October 1, 2026, which cites the sources named in brackets. We have not re-verified each original, so treat the list as a reported record.

• Eliezer Yudkowsky: above 95% (Fast Company, 2023).
• Roman Yampolskiy: 99.999999% (X and Business Insider, 2024).
• Max Tegmark: above 90% (X, 2025).
• Daniel Kokotajlo: 70% to 80% (YouTube, 2025).
• Paul Christiano: 50% (The Independent, 2023).
• Yoshua Bengio: 20%, "around, like, 20 per cent probability that it turns out catastrophic" (ABC News, 2023).
• Elon Musk: 10% to 30%, "20% chance AI destroys humanity" (Business Insider, 2024).
• Geoffrey Hinton: 10% to 20% (2024), and 10% in a BBC Newsnight interview reported in September 2026.
• Lex Fridman: about 10% (2025).
• Scott Aaronson: about 2% (Shtetl-Optimized, 2023).
• Yann LeCun: below 0.01%, "less likely than an asteroid wiping us out" (TechRadar, 2024).
• Marc Andreessen: 0% (The New Yorker, 2024).

What the lab chiefs said

Dario Amodei, the Anthropic chief executive, was asked for his number at the Axios AI+ DC Summit in September 2025 and answered: "I think there's a 25% chance that things go really, really badly." He added a 75% chance that things go really, really well, and said he did not like the term. Sam Altman and Demis Hassabis are listed by Wikipedia only as "above zero." In May 2023 all three, with Hinton and Bengio, signed a one-sentence Center for AI Safety statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."

Signing a statement is not a probability, and it is worth keeping the two apart. The statement says the risk deserves priority. It does not say the risk is likely.

Why the numbers are so far apart

A 2025 survey of 111 experts by Severin Field explains part of it. Experts clustered into two viewpoints, "AI as controllable tool" and "AI as uncontrollable agent." Only 21% had met the concept of instrumental convergence, though 78% thought researchers should address catastrophic risks. People who hold different premises about whether future systems will pursue goals of their own reach different totals.

A second reason is incentive. Lab leaders have commercial reasons to say AI is powerful and safety researchers have career reasons to say it is dangerous. Skeptics have the reverse incentives. That does not make any of them wrong, but it is a reason to ask what a person would stand to gain.

What the numbers cannot tell you

A p doom figure does not say when. A 10% chance of catastrophe by 2100 and a 10% chance within a decade are very different claims, and the headlines usually drop the horizon. It also does not say what could be done about it: a person can hold a high number and still think the right response is more safety research, a treaty, or a pause, and those views split further apart than the numbers do. When you see a figure quoted without a date, a definition and a horizon, treat it as a mood rather than a forecast.

How to read a p doom number

Check four things before you repeat one:

• The question. Extinction, loss of control, or any catastrophe?
• The date. Estimates move. Hinton's range has sat between 10% and 20% for at least two years.
• The type of source. A survey of thousands of researchers carries more weight than a podcast answer.
• The reasoning. Is there a stated mechanism, or only a number?

Nothing in this list has a track record, because nobody has resolved one of these forecasts. A calibrated forecaster needs outcomes to check against.

Our take

Use p doom as a signal of how worried a person is, never as evidence of how risky the technology is. The survey figure is the only number that tells you something about the field, and it says a large minority of researchers rate extinction-level risk at 10% or more. The honest consensus is that nobody knows, and that the uncertainty is itself a reason for testing and disclosure. We will add new estimates here as they are published, with the date.