P(doom): What It Means and Every Major AI Researcher’s Estimate

P(doom) is the estimated chance AI causes human extinction. See named researchers’ stated p(doom) with dates and sources, and why the numbers differ.

By Izhar Ahmad Danish · · 5 min read

P(doom): What It Means and Every Major AI Researcher’s Estimate: Time of Superintelligence cover image
On this page
  1. What does p(doom) mean?
  2. Researchers’ p(doom) estimates: the table
  3. What is Geoffrey Hinton’s p(doom)?
  4. Why do p(doom) estimates differ so much?
  5. Is p(doom) a useful number?
  6. Our take
  7. Frequently asked questions
  8. Sources

P(doom), short for “probability of doom,” is the number AI researchers give when asked how likely it is that artificial intelligence ends in catastrophe for humanity. It started as an in-joke in AI safety circles and is now a standard question in interviews with lab CEOs.

Short answer: P(doom) is a person’s estimated probability that AI causes human extinction or a similarly permanent disaster, such as humanity losing control to machines. Stated values range from under 0.01% (Yann LeCun) to over 99% (Roman Yampolskiy). Geoffrey Hinton says 10–20% and Anthropic’s Dario Amodei 25%. The 2023 survey median among AI researchers was 5%.

What does p(doom) mean?

Short answer: the probability, in a person’s own judgment, that AI leads to an existential catastrophe.

The “p” is probability notation, as in P(A). “Doom” usually means one of three things:

  • Human extinction caused by AI.
  • Permanent disempowerment: humans survive but lose control of their future to AI systems.
  • An irreversible catastrophe short of extinction, depending on who is speaking.

According to Wikipedia’s history of the term, it was first used around 2009 on the rationalist forum LessWrong and spread widely in 2023, after GPT-4 and after Geoffrey Hinton left Google to warn about AI. The same year, hundreds of researchers signed the Center for AI Safety’s one-sentence statement that “mitigating the risk of extinction from AI should be a global priority.”

Researchers’ p(doom) estimates: the table

This table lists stated figures from named people, the date they gave them and where. Each figure answers a slightly different question, so read the “what it covers” column before comparing.

Person Role P(doom) Date What it covers Source
Yann LeCun Turing Award winner; founder, AMI Labs <0.01% 2024 Calls AI doom less likely than an asteroid strike TechRadar
AI researchers (survey median) 2,778 published authors 5% 2023 survey, pub. Jan 2024 Extinction or similarly severe disempowerment Grace et al.
Geoffrey Hinton Nobel laureate; “godfather of AI” 10–20% 2024 Extinction or loss of control within ~30 years BBC Newsnight via Digital Today
Elon Musk CEO, xAI / SpaceX 10–20% Oct 2024 Chance AI “goes bad” Fortune
Yoshua Bengio Turing Award winner; LawZero ~20% Jul 2023 Catastrophic outcome; based partly on 50% odds of human-level AI within a decade Live Science
Dario Amodei CEO, Anthropic 25% Sep 2025 Chance things go “really, really badly” Axios
Paul Christiano Former OpenAI alignment lead ~50% 2023 “Doom shortly after” human-level AI; also 10–20% for AI takeover with most humans dead. He calls these rough guesses Fortune, LessWrong
Roman Yampolskiy AI safety researcher, Univ. of Louisville 99.9%+ 2024 Over the next 100 years TechRadar

Wikipedia’s P(doom) article maintains a longer list (including Eliezer Yudkowsky above 95%, Max Tegmark above 90% and Marc Andreessen at 0%). We include only figures we could trace to a dated report.

What the table shows

  • The spread is enormous. Two Turing Award co-winners, LeCun and Bengio, differ by a factor of more than 2,000.
  • Lab leaders are not optimists. The CEO of Anthropic gives 25% and Musk 10–20% while both keep building frontier models. Their argument is that safety-focused labs should be at the frontier rather than leave it to others.
  • The survey median is lower than the headline names. Most working researchers give single-digit odds, but a 5% median for extinction is still high for any technology. In the same survey, between 38% and 51% of respondents gave at least a 10% chance to outcomes as bad as extinction.

What is Geoffrey Hinton’s p(doom)?

Short answer: 10–20%, over roughly the next 30 years.

Hinton told BBC Newsnight that a 10% chance AI wipes out humanity is “not unreasonable,” and he has since repeated a 10–20% range. He frames the number as a gut figure that sits between optimists and the most worried researchers, not a calculation. The background to his view is on our godfather of AI page.

Why do p(doom) estimates differ so much?

Short answer: people disagree about three separate questions, and small differences multiply.

A p(doom) figure is usually the product of several guesses:

  1. Will we build superintelligence at all, and when? Bengio’s 20% assumed a 50% chance of human-level AI within a decade. Skeptics like LeCun think current methods won’t get there.
  2. Will it be misaligned? That is, will its goals differ from ours in ways we can’t correct? This is the AI alignment problem.
  3. Will a misaligned system be able to cause catastrophe, and will we fail to stop it?

If you think each step is 50/50, you land near 12%. If you think each is 90% likely, you land above 70%. That is why the same evidence produces such different numbers.

The time frame and the definition also vary. Yampolskiy’s 99.9% covers 100 years; Hinton’s covers about 30. Amodei’s 25% is “things go really badly,” which is broader than extinction.

Is p(doom) a useful number?

Short answer: as a measure of concern, yes; as a forecast, not really.

Critics, including Amodei himself, dislike the term because it turns deep uncertainty into a single number that sounds precise. There is no track record to calibrate against: humanity has never built superintelligence before.

Defenders say the numbers still matter. Asking for a probability forces people to be specific, and it exposes the gap between what lab leaders say privately and how fast they build. A 10% chance of catastrophe from any other product, such as an aircraft or a drug, would stop it from being approved.

Our take

The most useful fact in the table is not any single number but that almost nobody serious says zero. Even researchers who think doom is unlikely tend to give a figure that would be unacceptable for any other technology. That is the core argument behind the Statement on Superintelligence, which asks for a pause on superintelligence until it can be built safely.

If you want to know how a takeover could actually happen and what experts say against it, read Will AI take over the world?. For the specific technical risks behind these numbers, see superintelligence risks and the alignment problem.

Enjoyed this? Get the SI Briefing

One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.

Frequently asked questions

What is p(doom)?

P(doom) is the probability someone assigns to AI causing human extinction or a similarly permanent catastrophe, such as humanity losing control of its future. It began as slang among AI safety researchers and became mainstream in 2023 after GPT-4.

What is Geoffrey Hinton’s p(doom)?

Hinton has said there is roughly a 10 to 20 percent chance that AI leads to human extinction or a permanent loss of control within about 30 years. He stresses that nobody really knows.

What is the risk of AI causing human extinction?

Nobody knows. A 2023 survey of 2,778 published AI researchers gave a median of 5 percent for outcomes as bad as human extinction, while individual estimates range from near zero to over 99 percent.

What is Elon Musk’s p(doom)?

Musk has repeatedly put it at about 10 to 20 percent, saying in October 2024 that there is a sub-chance of 10 to 20 percent that AI goes bad, while calling a good outcome most likely.

Who has the highest p(doom)?

Among well-known researchers, Roman Yampolskiy of the University of Louisville gives one of the highest figures, above 99.9 percent over the next 100 years. Eliezer Yudkowsky is also usually listed above 95 percent.

10 Sources

  1. P(doom) (Wikipedia)
  2. Grace et al., “Thousands of AI Authors on the Future of AI” (2024)
  3. Axios: Amodei puts p(doom) at 25% (17 Sep 2025)
  4. Digital Today: Hinton says 10% chance AI could end humanity “not unreasonable” (BBC Newsnight)
  5. Fortune: Elon Musk says there is a 10–20% chance AI “goes bad” (30 Oct 2024)
  6. Fortune: ex-OpenAI safety researcher Paul Christiano warns of AI doom (3 May 2023)
  7. Paul Christiano, “My views on doom” (LessWrong, 2023)
  8. TechRadar: Yampolskiy, Musk and LeCun on the odds AI ends humanity (2024)
  9. Live Science: Yoshua Bengio on catastrophic AI risk
  10. Center for AI Safety, Statement on AI Risk (2023)
Before you go

Super intelligence is moving fast.

Don’t miss what happens next

Get one short email a week with what changed and why it matters. No spam. Unsubscribe anytime.