P(doom) is a shorthand, borrowed from probability notation, for the subjective probability that a speaker assigns to existentially catastrophic outcomes from artificial intelligence — most often human extinction or permanent, severe disempowerment. The specific outcome referred to varies between speakers, but generally tracks the class of scenarios described under AI existential risk (Source: en.wikipedia.org). Because it compresses a set of contested empirical and definitional judgments into a single number, it functions in practice as a conversational marker of where a speaker sits in AI-risk debates rather than as a measured quantity.
Origin and diffusion
The term began as informal shorthand within the rationalist community and among AI researchers, and reached wider circulation in 2023 following the release of GPT-4, as figures including Geoffrey Hinton and Yoshua Bengio publicly raised concerns about AI risk (Source: en.wikipedia.org). Press coverage in 2023 treated the spread of the term as itself notable: ABC News described it as having "started as a dark in-joke" (Source: abc.net.au), and The New York Times covered its adoption in Silicon Valley in December 2023 (Source: nytimes.com). Fast Company ran an explainer in July 2023 framing it as a metric a reader could calculate for themselves (Source: fastcompany.com).
The usage the term displaced was not absent — Shane Legg gave a numeric estimate of AI-caused human extinction in a 2011 LessWrong Q&A, well before the label was in circulation (Source: lesswrong.com) — but the compact notation made the estimate a routine thing to state in interviews and podcasts.
Survey evidence
The most widely cited aggregate figure comes from the 2023 Expert Survey on Progress in AI, run by AI Impacts, which asked AI researchers to estimate the probability that future AI advances lead to human extinction or similarly severe and permanent disempowerment within 100 years. The mean response was 14.4% and the median 5% (Source: wiki.aiimpacts.org; reported in vox.com). The gap between mean and median indicates a right-skewed distribution: a minority of high estimates pulls the average well above the typical respondent.
General-population surveys return higher numbers than researcher surveys. A February 2025 US survey by Survey 160 asked respondents for the probability that AI development "will someday lead to the extinction of all human life" on a 0–100 scale, and reported a mean of 35.6% and a median of 20%, with 31% of respondents above 50%; the margin of error was 9.84 percentage points on a design-effect-adjusted basis (Source: survey160.com).
Range of published estimates
Individual estimates span close to the full probability interval. The distribution below is drawn from compilations of publicly stated figures; each is a point-in-time statement and several of the speakers have revised their numbers.
| Speaker | Stated P(doom) | Role at the time of statement |
|---|---|---|
| Yann LeCun | <0.01% | Then chief AI scientist, Meta |
| Marc Andreessen | 0% | Co-founder, Andreessen Horowitz |
| Sam Altman | >0% | CEO, OpenAI |
| Demis Hassabis | >0% | CEO, Google DeepMind |
| Casey Newton | 5% | Technology journalist |
| Shane Legg | c. 5–50% | Co-founder and chief AGI scientist, Google DeepMind |
| Nate Silver | 5–10% | Statistician, founder of FiveThirtyEight |
| Toby Ord | 10% | Philosopher, author of The Precipice |
| Lex Fridman | 10% | Computer scientist and podcast host |
| Vitalik Buterin | 12% | Co-founder, Ethereum |
| Lina Khan | c. 15% | Then chair, Federal Trade Commission |
| Geoffrey Hinton | 10–20% all-things-considered; >50% independent impression | 2024 Nobel laureate in Physics |
| Yoshua Bengio | 20% | Director, Mila; Turing Award winner |
| Dario Amodei | 10–25% | CEO, Anthropic |
| Elon Musk | c. 10–30% | CEO of X, Tesla and SpaceX |
| Jan Leike | 10–90% | Alignment researcher, Anthropic |
| Paul Christiano | 50% | Head of research, US AI Safety Institute |
| Holden Karnofsky | 50% | Executive director, Open Philanthropy |
| Emad Mostaque | 50% | Co-founder, Stability AI |
| Zvi Mowshowitz | 70% | Writer on artificial intelligence |
| Daniel Kokotajlo | 70–80% | Founder, AI Futures Project; formerly OpenAI |
| Dan Hendrycks | >80%, up from c. 20% two years earlier | Director, Center for AI Safety |
| Andrew Critch | 85% | Founder, Center for Applied Rationality |
| Connor Leahy | 90%+ | Co-founder, EleutherAI |
| Max Tegmark | >90% | Physicist; co-founder, Future of Life Institute |
| Eliezer Yudkowsky | >95% | Founder, MIRI |
| Roman Yampolskiy | 99.9%–99.999999% within 100 years | Computer scientist, University of Louisville |
(Source for the table: en.wikipedia.org, which cites the underlying interviews, podcasts and posts individually.)
Two speakers illustrate how differently the number is derived. Bengio described arriving at 20% from component judgments — roughly a 50% probability that AI reaches human-level capability within a decade, and a greater-than-50% probability that AI or humans would then turn the technology against humanity at scale. Grady Booch, by contrast, gave a figure he characterized as equivalent to "P(all the oxygen in my room spontaneously moving to a corner thereby suffocating me)" (Source: en.wikipedia.org).
A July 2026 aggregation by the site Calcuja of more than fifty publicly stated estimates reported a median of approximately 20% and noted that the standard deviation across estimates exceeded the median. It also reported that AI safety researchers give systematically higher figures than mainstream machine-learning researchers, and that estimates have risen on average since 2020. The compilation covers publicly stated estimates only and is self-described as descriptive rather than authoritative (Source: calcuja.com).
Criticism
The most common objection is that the number is underspecified. A given estimate may or may not be conditional on artificial general intelligence being built at all, may cover any time horizon, and may define "doom" as biological extinction, permanent disempowerment, civilizational collapse, or loss of human autonomy — and speakers using the same figure frequently mean different things by it (Source: en.wikipedia.org; nytimes.com). Isaac King argued in a January 2024 LessWrong post titled "Stop talking about p(doom)" that the compression obscures more than it conveys (Source: lesswrong.com). A commentary published by the American Enterprise Institute made a related argument that unconditional estimates are the problem, and that a stated probability is only interpretable alongside the conditionals it rests on (Source: aei.org).
A separate line of criticism treats the term's rhetorical function rather than its precision. In the account summarized on How MAGA learned to love AI safety — Nicky Woolf (Transformer, October 2025), investor Roger McNamee argued that existential-risk framing operates as fundraising rather than analysis, and located the route to catastrophe in mass unemployment and its political consequences rather than in a misaligned system.
Andrew Ng makes a related argument about motive from within the research field. In a Bloomberg TV interview on September 17, 2026 he called existential-risk warnings from researchers at the top model makers "much more science fiction than science" and said the industry had stoked catastrophe fears earlier in the AI boom in an apparent bid to bolster publicity and shape regulation (Source: bloomberg.com). The Next Web reports that in a December 2023 statement to a US Senate forum he put the odds of AI causing human extinction at roughly one in 10 million over a century (Source: thenextweb.com).
Use in policy argument
The term appears in AI-policy settings as a compact way to state a risk premise before arguing for a policy. The export-controls lecture summarized in Inside My AI Law & Policy Class 23: When Silicon Valley's Effective Altruists Meet Washington's Export Controls (Farahany / Hamilton, November 2025) opens with a p(doom) poll of the class keyed to a Yudkowsky talk — the class split between under and over 50%, with none below 20% — and pairs it with p(win) and p(utopia) framings to present export controls as instruments for improving the odds, an argument grounded in effective altruism.
Economic work has taken the quantity as a modeling input rather than a rhetorical one. Charles I. Jones argued in American Economic Review: Insights (2024) that even small existential risks from AI would justify substantial investment in safety and alignment research (DOI 10.1257/aeri.20230570). Jakub Growiec and Klaus Prettner published "The economics of p(doom): Scenarios of existential risk and economic growth in the age of transformative AI" in Economic Modelling in 2026 (DOI 10.1016/j.econmod.2026.107718).
The Center for AI Safety statement on AI risk is the collective-signature counterpart to the individual estimate: it asserts that mitigating extinction risk from AI should be a global priority without committing signatories to any particular probability.
Open questions
- Whether stated estimates are conditional on AGI being developed, and over what horizon, is not standardized across speakers, which limits comparison between the figures in the table above.
- Whether the upward drift in estimates reported by aggregators reflects changed beliefs, changed composition of who is asked, or changed willingness to state a number publicly is not established by the available surveys.
Relationships
- depends-on: AI Existential Risk — the risk hypothesis the number is a summary of.
- related: AGI Timelines — the capability-timing judgment most stated estimates are conditioned on.
- related: Superintelligence, AI Alignment
- related: Effective Altruism — the tradition in which the expected-value framing originates.
- supports: Statement on AI Risk (CAIS) — the collective statement of the same concern without a stated probability.
- related: Inside My AI Law & Policy Class 23: When Silicon Valley's Effective Altruists Meet Washington's Export Controls (Farahany / Hamilton, November 2025) — uses the framing to argue for export controls.
- contradicts: How MAGA learned to love AI safety — Nicky Woolf (Transformer, October 2025) — carries McNamee's argument that the framing serves fundraising rather than analysis.
- related: Eliezer Yudkowsky, Dan Hendrycks, Daniel Kokotajlo, Geoffrey Hinton, Yoshua Bengio, Yann LeCun, Dario Amodei