{ "lastUpdated": "2026-11-07", "definition": "p(doom) is informal shorthand for the probability someone assigns to advanced AI ending badly. It is not one quantity. The rows here answer at least five different questions, and the marker shape says which.", "outcomeClasses": [ { "id": "extinction", "marker": "circle", "label": "Human extinction", "desc": "The probability that AI kills everyone." }, { "id": "disempowerment", "marker": "diamond", "label": "Extinction or permanent disempowerment", "desc": "Extinction, or humanity permanently and irreversibly losing control of its own future — the survey wording, and the broader claim." }, { "id": "control", "marker": "triangle", "label": "Loss of human control", "desc": "Humans losing control of advanced AI. Not the same as extinction: an uncontrolled system need not kill everyone." }, { "id": "masscasualty", "marker": "pentagon", "label": "Over 10% of humanity dies within five years", "desc": "A global catastrophe as the Forecasting Research Institute defines it. Terrible, but not extinction, and not the same question as the other shapes." }, { "id": "catastrophe", "marker": "square", "label": "Catastrophic outcome, undefined", "desc": "“Things go really badly” — severity stated, but the endpoint left open." } ], "professionClasses": [ { "id": "researcher", "icon": "microscope", "label": "AI / ML researcher", "group": "Knowledge", "color": "#7c8ce8" }, { "id": "academic", "icon": "graduation-cap", "label": "Academic", "group": "Knowledge", "color": "#9fa9f2" }, { "id": "philosopher", "icon": "scroll", "label": "Philosopher", "group": "Knowledge", "color": "#6b7ad2" }, { "id": "executive", "icon": "building-2", "label": "Founder or executive", "group": "Industry", "color": "#e0a463" }, { "id": "engineer", "icon": "cpu", "label": "Engineer", "group": "Industry", "color": "#cd8b57" }, { "id": "investor", "icon": "trending-up", "label": "Investor", "group": "Industry", "color": "#eec481" }, { "id": "forecaster", "icon": "chart-spline", "label": "Forecaster or statistician", "group": "Observers", "color": "#4ec3a2" }, { "id": "journalist", "icon": "newspaper", "label": "Writer or journalist", "group": "Observers", "color": "#7ad6ba" }, { "id": "policy", "icon": "gavel", "label": "Regulator or policymaker", "group": "Civic", "color": "#e087a8" }, { "id": "advocate", "icon": "megaphone", "label": "Campaigner", "group": "Civic", "color": "#ee9ec0" }, { "id": "survey", "icon": "users", "label": "Surveyed population", "group": "Civic", "color": "#a9b5c7" } ], "sourcingRule": "A figure is plotted only if it can be traced to the person. Every row is graded by how it was checked: read in the published report, own words read at the source, reported by a named outlet, or cited through a compilation whose original we could not open. Nothing is inferred from the strength of someone's language.", "sources": [ "Grace, K. et al. (2026). Advanced AI according to 1,580 researchers: uncertain, unsafe, and sooner than we thought. AI Impacts.", "Grace, K. et al. (2024). Thousands of AI Authors on the Future of AI. arXiv:2401.02843.", "Forecasting Research Institute (2026). Longitudinal Expert AI Panel, Wave 9: Probability of a Global AI Catastrophe.", "Karger, E. et al. (2023). Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament. Forecasting Research Institute.", "Murray, C.J.L., Kanem, N. et al. (2026). The Lancet Commission on 21st century global threats to health. The Lancet.", "Wikipedia, P(doom), and PauseAI (pauseai.info/pdoom), used as pointers and graded as such", "Individual public statements, posts, interviews and transcripts as cited per entry" ], "surveys": [ { "name": "AI researchers (2024)", "role": "Expert Survey on Progress in AI · 1,489 answers from researchers at six leading venues", "op": "median", "low": 1, "high": 25, "mid": 10, "outcome": "disempowerment", "horizon": null, "conditional": null, "stated": "2026-09", "reference": "Grace et al., “Advanced AI according to 1,580 researchers,” AI Impacts, Sept 2026 (survey run Dec 2024)", "notes": "Pooled across three question framings answered by different people. Median 10%, mean 18.2%, middle half of answers 1% to 25%. 51% gave at least 1 in 10; 12% gave zero. The median was 5% in the previous survey, but the mean only rose from 16.5% to 18.2%, and the authors describe a slight increase. These are separate surveys rather than the same people tracked over time, and about 10% of those invited responded.", "prior": { "mid": 5, "stated": "2024-01", "note": "Median was 5% in the 2023 survey (Grace et al., Jan 2024). The mean rose from 16.5% to 18.2%." }, "ghost": { "value": 18.2, "op": "exact", "label": "mean" }, "profession": "survey", "evidence": "published", "url": "https://aiimpacts.org/wp-content/uploads/2026/09/ESPAI2024.pdf" }, { "name": "Superforecasters (2022)", "role": "Forecasting Research Institute · Existential Risk Persuasion Tournament", "op": "exact", "low": null, "high": null, "mid": 0.38, "outcome": "extinction", "horizon": "by 2100", "conditional": null, "stated": "2022", "reference": "Karger et al., Forecasting Existential Risks, Forecasting Research Institute, 2023 (tournament run Jun to Oct 2022), Table 9", "notes": "Median forecast for AI-caused extinction by 2100 (95% interval 0.10 to 0.75%), from accuracy-selected generalist forecasters. For AI killing over 10% of humanity within five years their median was 2.13%. The gap with the domain experts did not close when the two groups argued it out.", "prior": null, "ghost": null, "profession": "forecaster", "evidence": "published", "url": "https://forecastingresearch.org/pdf/existential-risk-persuasion-tournament.pdf" }, { "name": "AI experts (2022)", "role": "Forecasting Research Institute · Existential Risk Persuasion Tournament", "op": "exact", "low": null, "high": null, "mid": 3, "outcome": "extinction", "horizon": "by 2100", "conditional": null, "stated": "2022", "reference": "Karger et al., Forecasting Existential Risks, Forecasting Research Institute, 2023 (tournament run Jun to Oct 2022), Table 9", "notes": "Median forecast for AI-caused extinction by 2100 (95% interval 0.49 to 10%), from the domain experts who argued it out with the superforecasters. For AI killing over 10% of humanity within five years their median was 12%. The 2026 Lancet Commission used both groups' answers to put AI misuse among the two threats to health with extinction potential this century.", "prior": null, "ghost": null, "profession": "survey", "evidence": "published", "url": "https://forecastingresearch.org/pdf/existential-risk-persuasion-tournament.pdf" }, { "name": "Experts, rapid (LEAP 2026)", "role": "Forecasting Research Institute · 192 experts", "op": "median", "low": 3, "high": 22, "mid": 10, "outcome": "masscasualty", "horizon": "by 2100", "conditional": "if AI progress is rapid", "stated": "2026-06", "reference": "Forecasting Research Institute, LEAP Wave 9, forecasts collected 19 May to 10 Jun 2026", "url": "https://leap.forecastingresearch.org/forecasts/ai-catastrophic-risk", "notes": "A conditional forecast, not a headline figure: the same experts, asked to assume rapid AI progress. Under slow progress their median is 2%.", "prior": null, "ghost": null, "profession": "survey", "evidence": "published" }, { "name": "Public (LEAP 2026)", "role": "Forecasting Research Institute · 612 members of the public", "op": "median", "low": 3, "high": 14, "mid": 7, "outcome": "masscasualty", "horizon": "by 2100", "conditional": null, "stated": "2026-06", "reference": "Forecasting Research Institute, LEAP Wave 9, forecasts collected 19 May to 10 Jun 2026", "url": "https://leap.forecastingresearch.org/forecasts/ai-catastrophic-risk", "notes": "Higher than experts or superforecasters at every date, and rising to 15% under rapid progress.", "prior": null, "ghost": null, "profession": "survey", "evidence": "published" }, { "name": "Experts (LEAP 2026)", "role": "Forecasting Research Institute · 192 experts", "op": "median", "low": 1.5, "high": 10, "mid": 5, "outcome": "masscasualty", "horizon": "by 2100", "conditional": null, "stated": "2026-06", "reference": "Forecasting Research Institute, LEAP Wave 9, forecasts collected 19 May to 10 Jun 2026", "url": "https://leap.forecastingresearch.org/forecasts/ai-catastrophic-risk", "notes": "Computer scientists, industry professionals, economists and policy researchers. 0.3% by 2030, 2% by 2050, 5% by 2100. This is not the survey's extinction question: compare it only with the same shape.", "prior": null, "ghost": null, "profession": "survey", "evidence": "published" }, { "name": "Superforecasters (LEAP 2026)", "role": "Forecasting Research Institute · 53 superforecasters", "op": "median", "low": 0.5, "high": 6, "mid": 2.4, "outcome": "masscasualty", "horizon": "by 2100", "conditional": null, "stated": "2026-06", "reference": "Forecasting Research Institute, LEAP Wave 9, forecasts collected 19 May to 10 Jun 2026", "url": "https://leap.forecastingresearch.org/forecasts/ai-catastrophic-risk", "notes": "0.1% by 2030, 0.88% by 2050, 2.4% by 2100. Statistically distinct from the experts in the unconditional forecast at all three dates.", "prior": null, "ghost": null, "profession": "survey", "evidence": "published" } ], "estimates": [ { "name": "Roman Yampolskiy", "role": "AI safety researcher, University of Louisville", "op": "range", "low": 99.9, "high": 99.999999, "mid": 99.9, "outcome": "extinction", "horizon": "within the next 100 years", "conditional": null, "stated": "2024-06", "reference": "Business Insider, Jun 2024; @romanyam, 12 Mar 2024", "notes": "Argues the control problem is not solvable even in principle, so this is closer to a claim about mathematics than a forecast about events.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "reported", "url": "https://www.businessinsider.com/ai-researcher-roman-yampolskiy-lex-fridman-human-extinction-prediction-2024-6" }, { "name": "Eliezer Yudkowsky", "role": "Founder, Machine Intelligence Research Institute", "op": "gt", "low": null, "high": null, "mid": 95, "outcome": "extinction", "horizon": null, "conditional": "if superintelligence is built under present conditions", "stated": "2023-12-07", "reference": "Fast Company, 7 Dec 2023, as cited by Wikipedia", "notes": "Co-author of “If Anyone Builds It, Everyone Dies” (2025). Holds that a misaligned superintelligence kills everyone by default, and that nothing in current practice changes that default.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "cited", "url": "https://en.wikipedia.org/wiki/P(doom)" }, { "name": "Connor Leahy", "role": "Co-founder, EleutherAI; CEO, Conjecture", "op": "gt", "low": null, "high": null, "mid": 90, "outcome": "extinction", "horizon": null, "conditional": "conditional on AGI being built", "stated": "2024-11-22", "reference": "Future of Life Institute podcast, 22 Nov 2024, as cited by Wikipedia", "notes": "Co-founder of EleutherAI and chief executive of Conjecture, and a prominent public critic of the race to build AGI.", "prior": null, "ghost": null, "profession": "executive", "evidence": "cited", "url": "https://www.youtube.com/watch?v=nUV5-SLdkuQ" }, { "name": "Max Tegmark", "role": "Physicist, MIT; co-founder, Future of Life Institute", "op": "gt", "low": null, "high": null, "mid": 90, "outcome": "control", "horizon": null, "conditional": "if development continues on its current, largely unregulated trajectory", "stated": "2025-04-30", "reference": "@tegmark, “My P(doom) Estimate,” 30 Apr 2025, as cited by Wikipedia", "notes": "This is a probability of losing control, not of extinction — the two are often quoted interchangeably and should not be. In August 2026 he told a Cambridge audience the question is already settled: “This is in the rearview mirror. Clearly, multiple companies lost control over AI systems this summer.”", "prior": null, "ghost": null, "profession": "academic", "evidence": "cited", "url": "https://x.com/tegmark/status/1917580821101437280" }, { "name": "David Duvenaud", "role": "ML professor, University of Toronto; former Anthropic safety team lead", "op": "exact", "low": null, "high": null, "mid": 85, "outcome": "disempowerment", "horizon": null, "conditional": null, "stated": "2025-04-18", "reference": "Doom Debates, as cited by Wikipedia", "notes": "Counts the loss of everything he values, not only deaths — a world where humans survive but no longer decide anything is inside his 85%.", "prior": null, "ghost": null, "profession": "academic", "evidence": "cited", "url": "https://www.youtube.com/watch?v=mb9w7lFIHRM" }, { "name": "Andrew Critch", "role": "AI researcher, UC Berkeley", "op": "exact", "low": null, "high": null, "mid": 85, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2024-11-15", "reference": "Doom Debates, 15 Nov 2024, as cited by Wikipedia", "notes": "Focuses on multi-agent failure — not one system turning on us, but many competing systems squeezing humans out of the economy they run.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "cited", "url": "https://www.youtube.com/watch?v=opIvVzJF8t0" }, { "name": "Dan Hendrycks", "role": "Director, Center for AI Safety", "op": "gt", "low": null, "high": null, "mid": 80, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2023-12-07", "reference": "Fast Company, 7 Dec 2023, as cited by Wikipedia", "notes": "Organised the 2023 one-sentence statement on extinction risk signed by the heads of OpenAI, Google DeepMind and Anthropic.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "cited", "url": "https://en.wikipedia.org/wiki/P(doom)" }, { "name": "David Krueger", "role": "Assistant professor, Université de Montréal; Mila; founder, Evitable", "op": "exact", "low": null, "high": null, "mid": 75, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2026-09-02", "reference": "The Smart Cookies Podcast, as cited by Wikipedia", "notes": "A co-author of the 2026 expert survey report, with a number far above the median it found. In April 2026 an interviewer put it to him that his estimate was above 50%, and he did not dispute it.", "prior": null, "ghost": null, "profession": "academic", "evidence": "cited", "url": "https://www.youtube.com/watch?v=PyBw-dIcKW4" }, { "name": "Daniel Kokotajlo", "role": "Founder, AI Futures Project; formerly OpenAI", "op": "exact", "low": null, "high": null, "mid": 70, "outcome": "disempowerment", "horizon": null, "conditional": "if things do not change", "stated": "2026-10-07", "reference": "Substack interview, “The Anteaters Can’t Do Anything,” 7 Oct 2026", "notes": "“I usually say my p(doom) is 70%, but that's not necessarily human extinction.” Elsewhere: AIs taking over, or a very big catastrophe like that. He gave 70 to 80% in April 2025.", "prior": null, "ghost": null, "profession": "forecaster", "evidence": "own", "url": "https://pogueman.substack.com/p/the-anteaters-cant-do-anything-daniel" }, { "name": "Zvi Mowshowitz", "role": "AI writer, Don't Worry About the Vase", "op": "exact", "low": null, "high": null, "mid": 70, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-09-06", "reference": "The Cognitive Revolution, 6 Sep 2025, as cited by Wikipedia", "notes": "Revised upward, saying the bad news had outweighed the good — though not by enough to reach 80%.", "prior": { "mid": 60, "stated": "2023", "note": "Previously 60%." }, "ghost": null, "profession": "journalist", "evidence": "cited" }, { "name": "Holden Karnofsky", "role": "Anthropic; co-founder of Open Philanthropy, now Coefficient Giving", "op": "range", "low": 10, "high": 90, "mid": 50, "outcome": "control", "horizon": null, "conditional": "if transformative AI arrives soon", "stated": "2022-08-29", "reference": "“How might we align transformative AI if it’s developed very soon?”, Alignment Forum, 29 Aug 2022", "notes": "A range for AI takeover, not a central figure: he doubts anyone should go under 10% or over 90%. The 50% often attributed to him comes from compilations.", "prior": null, "ghost": null, "profession": "executive", "evidence": "own", "url": "https://www.alignmentforum.org/posts/rCJQAkPTEypGjSJ8X/how-might-we-align-transformative-ai-if-it-s-developed-very" }, { "name": "Emad Mostaque", "role": "Co-founder, Stability AI", "op": "exact", "low": null, "high": null, "mid": 50, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2024-12-04", "reference": "@EMostaque, 4 Dec 2024", "notes": "“Given an undefined time period the probability of systems that are more capable than humans and likely end up running all our critical infrastructure wiping us all out is a coin toss.” A blog reports he said 20% in September 2026; we have not found his own words.", "prior": null, "ghost": null, "profession": "executive", "evidence": "own", "url": "https://x.com/EMostaque/status/1864266899170767105" }, { "name": "Jan Leike", "role": "Alignment researcher, Anthropic; formerly OpenAI, DeepMind", "op": "range", "low": 10, "high": 90, "mid": 50, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023-08-07", "reference": "80,000 Hours, 7 Aug 2023", "notes": "An endorsed range, not a forecast of his own. He called the question not very useful; when the interviewer guessed more than 10% and less than 90%, he said “That's probably the range I would give too.”", "prior": null, "ghost": null, "profession": "researcher", "evidence": "own", "url": "https://80000hours.org/podcast/episodes/jan-leike-superalignment/" }, { "name": "Paul Christiano", "role": "OpenAI Foundation board; formerly head of AI safety, US AI Safety Institute", "op": "exact", "low": null, "high": null, "mid": 46, "outcome": "disempowerment", "horizon": "within 10 years of building powerful AI", "conditional": null, "stated": "2023-04", "reference": "LessWrong, “My views on doom,” Apr 2023", "notes": "Decomposed rather than guessed: roughly 22% for an outright misalignment disaster, the rest for other routes to the same end. The inventor of RLHF, the technique that made modern chatbots usable.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "own", "url": "https://www.lesswrong.com/posts/xWMqsvHapP3nwdSW8/my-views-on-doom" }, { "name": "Joep Meindertsma", "role": "Founder, PauseAI", "op": "exact", "low": null, "high": null, "mid": 40, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2024", "reference": "pauseai.info, his own organisation's site; NRC interview, 2024", "notes": "“The remaining 60% consists largely of: we can pause.”", "prior": null, "ghost": null, "profession": "advocate", "evidence": "cited", "url": "https://pauseai.info/pdoom" }, { "name": "Scott Alexander", "role": "Writer, Astral Codex Ten", "op": "exact", "low": null, "high": null, "mid": 33, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023", "reference": "Astral Codex Ten, 2023", "notes": "Later said he had considered revising down toward 20%, without committing to it.", "prior": null, "ghost": null, "profession": "journalist", "evidence": "own", "url": "https://www.astralcodexten.com/p/why-i-am-not-as-much-of-a-doomer" }, { "name": "Elon Musk", "role": "CEO, Tesla and SpaceX, which absorbed xAI", "op": "exact", "low": null, "high": null, "mid": 20, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2025", "reference": "The Joe Rogan Experience, 2025, reported by Business Insider", "notes": "“The probability of a good outcome is like 80%.” In March 2024 he said “about 10% or 20% or something like that.” He has argued the technology should be built anyway.", "prior": null, "ghost": null, "profession": "executive", "evidence": "reported", "url": "https://www.aol.com/elon-musk-says-theres-only-001804145.html" }, { "name": "Yoshua Bengio", "role": "Co-president and scientific director, LawZero; founder, Mila; 2018 Turing Award", "op": "exact", "low": null, "high": null, "mid": 20, "outcome": "catastrophe", "horizon": null, "conditional": "from ~50% odds on human-level AI within a decade and better-than-even odds of misuse or misalignment", "stated": "2023-07-14", "reference": "ABC News, 14 Jul 2023", "notes": "“I got around, like, 20 per cent probability that it turns out catastrophic.” Built from two component judgements rather than stated whole.", "prior": null, "ghost": null, "profession": "academic", "evidence": "reported", "url": "https://www.abc.net.au/news/2023-07-15/whats-your-pdoom-ai-researchers-worry-catastrophe/102591340" }, { "name": "Dario Amodei", "role": "CEO and co-founder, Anthropic", "op": "exact", "low": null, "high": null, "mid": 25, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-09-17", "reference": "Axios AI+ DC Summit, 17 Sep 2025, reported by Axios and TechRadar", "notes": "“I think there's a 25% chance that things go really, really badly.” He includes loss of control, national-security disasters and socioeconomic collapse in the same figure, which is why it is not an extinction number. In 2023 he gave a range of 10 to 25%.", "prior": null, "ghost": null, "profession": "executive", "evidence": "reported", "url": "https://www.axios.com/2025/09/17/anthropic-dario-amodei-p-doom-25-percent" }, { "name": "Geoffrey Hinton", "role": "2024 Nobel laureate in Physics; formerly Google", "op": "range", "low": 10, "high": 20, "mid": 15, "outcome": "extinction", "horizon": "in the coming decades", "conditional": null, "stated": "2026-10-13", "reference": "The Guardian, 13 Oct 2026", "notes": "“The point of giving numbers is to say this is a serious risk. We don't know how serious, but it's serious.” To the BBC, September 2026: “nobody really knows how to give a sensible estimate.” The 2024 split between 10 to 20% and above 50% is as listed by Wikipedia from a METR Q&A we have not heard.", "prior": null, "ghost": { "value": 50, "op": "gt", "label": "independent impression, 2024, as cited by Wikipedia" }, "profession": "academic", "evidence": "reported", "url": "https://www.theguardian.com/us-news/2026/oct/13/geoffrey-hinton-trump-ai-tech" }, { "name": "Lina Khan", "role": "Former chair, US Federal Trade Commission", "op": "approx", "low": null, "high": null, "mid": 15, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023-11", "reference": "Hard Fork, Nov 2023, as reported in compilations", "notes": "The only regulator here to give a number.", "prior": null, "ghost": null, "profession": "policy", "evidence": "cited", "url": "https://en.wikipedia.org/wiki/P(doom)" }, { "name": "Vitalik Buterin", "role": "Co-founder, Ethereum", "op": "exact", "low": null, "high": null, "mid": 12, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-08-12", "reference": "Doom Debates, 12 Aug 2025, as cited by Wikipedia", "notes": "Argues for “d/acc” — accelerating defensive and decentralising technology specifically so that the number comes down.", "prior": null, "ghost": null, "profession": "engineer", "evidence": "cited", "url": "https://www.youtube.com/watch?v=hxlcQzvnmWI&t=555s" }, { "name": "Evan Hubinger", "role": "Alignment science lead, Anthropic", "op": "gt", "low": null, "high": null, "mid": 10, "outcome": "extinction", "horizon": "within the next decade", "conditional": null, "stated": "2026-09-08", "reference": "@EvanHub, 8 Sep 2026", "notes": "“Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” Added that Anthropic has no plan for aligning superintelligence and is not clearly on track to find one, and that his figure is on the optimistic side among people at frontier labs.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "own", "url": "https://x.com/EvanHub/status/2097497037956891126" }, { "name": "Joe Carlsmith", "role": "Anthropic; formerly Open Philanthropy", "op": "gt", "low": null, "high": null, "mid": 10, "outcome": "disempowerment", "horizon": "by 2070", "conditional": null, "stated": "2023-03", "reference": "joecarlsmith.com, “Existential Risk from Power-Seeking AI”", "notes": "The most explicitly worked-out estimate on this chart: a six-step conjunctive model, each step assigned a probability and multiplied out.", "prior": { "mid": 5, "stated": "2021", "note": "His original report put it near 5%; he roughly doubled it a year later." }, "ghost": null, "profession": "philosopher", "evidence": "own", "url": "https://joecarlsmith.com/2023/03/22/existential-risk-from-power-seeking-ai-shorter-version" }, { "name": "Toby Ord", "role": "Philosopher, Oxford; author of The Precipice", "op": "exact", "low": null, "high": null, "mid": 10, "outcome": "disempowerment", "horizon": "this century", "conditional": null, "stated": "2020", "reference": "ABC News, Oct 2023; The Precipice (2020)", "notes": "Identifies unaligned AI as the largest single existential risk this century — larger than nuclear war, pandemics or climate change in his accounting. Said in 2024 that his figures had not moved enough to revise.", "prior": null, "ghost": null, "profession": "philosopher", "evidence": "reported", "url": "https://www.abc.net.au/news/2023-10-08/is-there-really-a-1-in-6-chance-of-human-extinction/102942530" }, { "name": "Lex Fridman", "role": "Research scientist, MIT; podcast host", "op": "exact", "low": null, "high": null, "mid": 10, "outcome": "extinction", "horizon": null, "conditional": "conditional on AGI leading to superintelligence", "stated": "2025-06-05", "reference": "Lex Fridman Podcast, Sundar Pichai episode, 5 Jun 2025", "notes": "“I would say my p(doom) is about 10%.”", "prior": null, "ghost": null, "profession": "journalist", "evidence": "own", "url": "https://lexfridman.com/sundar-pichai-transcript/" }, { "name": "Emmett Shear", "role": "Co-founder, Twitch; former interim CEO, OpenAI", "op": "range", "low": 5, "high": 50, "mid": 27.5, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023", "reference": "Logan Bartlett interview, reported by Futurism", "notes": "Ran OpenAI for three days during the November 2023 board crisis.", "prior": null, "ghost": null, "profession": "executive", "evidence": "reported", "url": "https://futurism.com/the-byte/openai-interim-ceo-ai-doomer" }, { "name": "Shane Legg", "role": "Co-founder and chief AGI scientist, Google DeepMind", "op": "range", "low": 5, "high": 50, "mid": 27.5, "outcome": "extinction", "horizon": "within about a year of human-level AI arriving", "conditional": "conditional on badly done AI", "stated": "2011-05-17", "reference": "LessWrong Q&A, 17 May 2011", "notes": "The oldest estimate on this chart by more than a decade — given in 2011, before deep learning worked. He has not publicly revised it, which is itself worth knowing.", "prior": null, "ghost": null, "profession": "researcher", "evidence": "own", "url": "https://www.lesswrong.com/posts/No5JpRCHzBrWA4jmS/q-and-a-with-shane-legg-on-risks-from-ai" }, { "name": "Nate Silver", "role": "Statistician; founder, FiveThirtyEight", "op": "range", "low": 5, "high": 10, "mid": 7.5, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-01-27", "reference": "Silver Bulletin, 27 Jan 2025", "notes": "“My estimate of p(doom)… is well in line with the expert consensus, which puts the chances at 5 to 10 percent.” A soft statement, and his definition includes dystopia as well as deaths.", "prior": null, "ghost": null, "profession": "forecaster", "evidence": "own", "url": "https://www.natesilver.net/p/its-time-to-come-to-grips-with-ai" }, { "name": "Benjamin Mann", "role": "Co-founder, Anthropic", "op": "range", "low": 0, "high": 10, "mid": 5, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-07-20", "reference": "Lenny's Podcast, 20 Jul 2025, as cited by Wikipedia", "notes": "Among the lowest figures given by anyone inside a frontier lab.", "prior": null, "ghost": null, "profession": "executive", "evidence": "cited", "url": "https://www.youtube.com/watch?v=WWoyWNhx2XU" }, { "name": "Kevin Roose", "role": "Technology columnist, The New York Times", "op": "range", "low": 5, "high": 15, "mid": 10, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023-12-07", "reference": "Fast Company, 7 Dec 2023, as cited by Wikipedia", "notes": "One of the journalists who has spent the most time with these systems.", "prior": null, "ghost": null, "profession": "journalist", "evidence": "cited", "url": "https://en.wikipedia.org/wiki/P(doom)" }, { "name": "Casey Newton", "role": "Founder, Platformer", "op": "exact", "low": null, "high": null, "mid": 5, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2023-12-07", "reference": "Fast Company, 7 Dec 2023, as cited by Wikipedia", "notes": "Covers the industry daily and sits an order of magnitude below the researchers he reports on.", "prior": null, "ghost": null, "profession": "journalist", "evidence": "cited", "url": "https://en.wikipedia.org/wiki/P(doom)" }, { "name": "Gary Marcus", "role": "Cognitive scientist; professor emeritus, NYU", "op": "approx", "low": null, "high": null, "mid": 3, "outcome": "catastrophe", "horizon": null, "conditional": null, "stated": "2025-07", "reference": "Marcus on AI, “Why my p(doom) has risen, dramatically,” Jul 2025", "notes": "One of the loudest sceptics of current AI capability claims — and he raised his number sharply in 2025. His scenario is not an autonomous machine but a reckless person with a powerful one.", "prior": null, "ghost": null, "profession": "academic", "evidence": "own", "url": "https://garymarcus.substack.com/p/why-my-pdoom-has-risen-dramatically" }, { "name": "William MacAskill", "role": "Philosopher, Oxford; author of What We Owe the Future", "op": "range", "low": 1, "high": 10, "mid": 5.5, "outcome": "disempowerment", "horizon": null, "conditional": null, "stated": "2025", "reference": "80,000 Hours, “Century in a decade,” 2025", "notes": "“I'm still on the more optimistic end: 1% to 10%.” He places himself at the hopeful end of a community largely defined by this worry.", "prior": null, "ghost": null, "profession": "philosopher", "evidence": "own", "url": "https://80000hours.org/podcast/episodes/will-macaskill-century-in-a-decade-navigating-intelligence-explosion/" }, { "name": "Scott Aaronson", "role": "Theoretical computer scientist, UT Austin; formerly OpenAI", "op": "approx", "low": null, "high": null, "mid": 2, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2023", "reference": "Shtetl-Optimized, 2023", "notes": "“A ~2% probability for the classic AGI-doom paperclip-maximizer-like scenario”, with a much higher figure for catastrophes in which AI is causally involved. He treats it lightly. Wikipedia now lists 5 to 50% from a 2026 interview we have not heard.", "prior": null, "ghost": null, "profession": "academic", "evidence": "own", "url": "https://scottaaronson.blog/?p=7672" }, { "name": "Geoffrey Irving", "role": "Formerly chief scientist, UK AI Security Institute; formerly OpenAI", "op": "approx", "low": null, "high": null, "mid": 50, "outcome": "extinction", "horizon": null, "conditional": null, "stated": "2026-10-15", "reference": "@geoffreyirving, 15 Oct 2026", "url": "https://x.com/geoffreyirving/status/2110761297688346776", "notes": "“In my view we have about a coin-flip chance that everyone dies as a result of superintelligence AI.”", "prior": null, "ghost": null, "profession": "researcher", "evidence": "own" } ], "unnumbered": [ { "name": "Jensen Huang", "role": "CEO, NVIDIA", "position": "“2030 is not going to be the end of the world… There is a 0% chance it's going to be the end of the world.”", "stated": "2026-09-18", "reference": "CBS, reported by Bloomberg, 18 Sep 2026", "note": "Time-bounded to 2030, so not a claim of zero risk forever. Zero cannot be drawn on a logarithmic axis. The exact question put to him is not in the reports we read.", "profession": "executive", "evidence": "reported" }, { "name": "Yann LeCun", "role": "Formerly chief AI scientist, Meta; 2018 Turing Award", "position": "Not worried “at all” about AI wiping out humanity.", "stated": "2026-10-01", "reference": "Fortune, 1 Oct 2026", "note": "The <0.01% in compilations traces to a third party's post and an unquoted line in TechRadar, not to a figure we can find in his own words, so it has no dot.", "profession": "researcher", "evidence": "reported" }, { "name": "Jacob Coxon", "role": "Former researcher, Anthropic and OpenAI", "position": "“The people building AI earnestly believe that it could kill us all by the end of the decade.”", "stated": "2026-09-08", "reference": "Resignation post on X, 8 Sep 2026", "note": "The resignation that made this argument public, seen over 115 million times. He gave no probability — the >10% belongs to Hubinger, who replied to him.", "profession": "researcher", "evidence": "reported" }, { "name": "Sam Altman", "role": "CEO, OpenAI", "position": "Greater than zero.", "stated": "2025-10-01", "reference": "MD MEETS, episode 1, 1 Oct 2025, as cited by Wikipedia", "note": "Has described catastrophic misalignment as “lights out for all of us” but has consistently declined to attach a figure.", "profession": "executive", "evidence": "cited" }, { "name": "Demis Hassabis", "role": "CEO, Google DeepMind; 2024 Nobel laureate in Chemistry", "position": "Greater than zero — “definitely non-zero and probably non-negligible.”", "stated": "2024-02-23", "reference": "Hard Fork, 23 Feb 2024, as cited by Wikipedia", "note": "“I don't know how people are getting their p(doom) numbers.” Declines the exercise while calling the risk serious enough to warrant international coordination.", "profession": "executive", "evidence": "cited" }, { "name": "Nate Soares", "role": "President, Machine Intelligence Research Institute", "position": "The title of his 2025 book with Eliezer Yudkowsky: If Anyone Builds It, Everyone Dies.", "stated": "2025-09", "reference": "If Anyone Builds It, Everyone Dies (2025)", "note": "No public numerical estimate we could source to him directly, despite a position as strong as any on the chart.", "profession": "researcher", "evidence": "own" }, { "name": "Grady Booch", "role": "Software engineer; IBM Fellow", "position": "About zero.", "stated": "2023-12-07", "reference": "Fast Company, 7 Dec 2023, as cited by Wikipedia", "note": "Compares the probability to all the oxygen in a room spontaneously moving into one corner.", "profession": "engineer", "evidence": "cited" }, { "name": "Marc Andreessen", "role": "Co-founder, Andreessen Horowitz", "position": "“p(Doom) = 0”, added to his bio on X.", "stated": "2024-03", "reference": "The New Yorker, Mar 2024", "note": "Not a small number but the number zero. It cannot be drawn on a logarithmic axis, which is part of why it belongs here rather than on the chart.", "profession": "investor", "evidence": "reported" }, { "name": "Reid Hoffman", "role": "Co-founder, LinkedIn", "position": "“I'm going to go for two on that answer”, on a scale from 1, no threat, to 10, extinction.", "stated": "2024", "reference": "PBS NewsHour", "note": "A rating on a scale, not a probability. Compilations often turn it into 20%.", "profession": "investor", "evidence": "reported" } ], "thresholds": [ { "name": "Asteroid Warning", "kind": "Action trigger", "low": 1, "high": null, "mid": 1, "display": "1 in 100 chance of impact", "reference": "International Asteroid Warning Network and Space Mission Planning Advisory Group, Thresholds and Criteria, 2018", "url": "https://cosmos.esa.int/documents/336356/1879207/SMPAG-RP-003_01_0_Thresholds%26Criterion_2018-10-18.pdf", "notes": "Above 1%, the international network must warn for any object over 10 metres, and mission planning starts for objects over 50 metres due within 50 years. Preparedness on the ground starts above 10%. 2024 YR4 was the first object to pass the 1% line. It is a line for acting, not a limit on what is tolerated.", "evidence": "published" }, { "name": "Dam Safety", "kind": "Safety limit", "low": 0.01, "high": null, "mid": 0.01, "display": "1 in 10,000 a year", "reference": "FERC Risk-Informed Decision Making Guidelines, ch. 3; US Bureau of Reclamation, Best Practices A-9", "url": "https://ferc.gov/sites/default/files/2020-04/chapter-3.pdf", "notes": "A limit on the yearly chance of one dam failing, or on the risk to the most exposed person. Above it, risk is unacceptable except in extraordinary circumstances. It is a standard for one structure, not a tolerance for every risk.", "evidence": "published" }, { "name": "Nuclear Reactors", "kind": "Regulatory benchmark", "low": 0.01, "high": null, "mid": 0.01, "display": "1 in 10,000 per reactor-year", "reference": "US NRC Safety Goal Policy; Regulatory Guide 1.174", "url": "https://www.nrc.gov/docs/ML0118/ML011800133.pdf", "notes": "A benchmark for core damage, meaning damage to the fuel in one reactor, with 1 in 100,000 for a large early release. It measures an accident at one plant, not deaths.", "evidence": "reported" }, { "name": "Ocean Worlds", "kind": "Safety limit", "low": 0.01, "high": null, "mid": 0.01, "display": "1 in 10,000 per mission", "reference": "COSPAR Policy on Planetary Protection, 2021", "url": "https://cosparhq.cnes.fr/assets/uploads/2021/07/PPPolicy_2021_3-June.pdf", "notes": "The most a mission may risk contaminating the subsurface oceans of Europa or Enceladus with Earth life. One of the few limits written to protect a whole world.", "evidence": "published" }, { "name": "Public Risk (UK)", "kind": "Safety limit", "low": 0.0001, "high": 0.01, "mid": 0.01, "display": "1 in 10,000 a year, tolerable limit for the public; 1 in a million, broadly acceptable", "reference": "UK Health and Safety Executive, Reducing Risks, Protecting People, 2001", "url": "https://assets.publishing.service.gov.uk/media/6634988fcf3b5081b14f30cd/IQ8.10.J_Document_9_Health_and_Safety_Executive__Reducing_risks__protecting_people__HSE_s_decision-making_process__2001.pdf", "notes": "The risk of death to one person from a hazardous activity. For workers the tolerable limit is 1 in 1,000 a year.", "evidence": "published" }, { "name": "Drinking Water", "kind": "Regulatory benchmark", "low": 0.0001, "high": 0.01, "mid": 0.001, "display": "1 in 100,000 over a lifetime (WHO); 1 in a million to 1 in 10,000 (US EPA)", "reference": "WHO Guidelines for Drinking-water Quality and US EPA practice, as summarised by state and national agencies", "notes": "Extra lifetime cancer risk from a carcinogen in the water supply. WHO itself calls the choice of level a value judgment.", "evidence": "reported" }, { "name": "Medicines", "kind": "Safety limit", "low": 0.001, "high": null, "mid": 0.001, "display": "1 in 100,000 over a lifetime", "reference": "ICH M7(R2), Assessment and Control of DNA Reactive Impurities, 2023", "url": "https://database.ich.org/sites/default/files/ICH_M7%28R2%29_Guideline_Step4_2023_0216_0.pdf", "notes": "The theoretical extra cancer risk accepted from a mutagenic impurity in a drug, which sets the usual limit of 1.5 micrograms a day. It is not a limit on the drug's own side effects.", "evidence": "published" }, { "name": "Hazardous Sites (NL)", "kind": "Safety limit", "low": 0.0001, "high": null, "mid": 0.0001, "display": "1 in a million a year", "reference": "Dutch External Safety Decree (Bevi), as applied in municipal planning documents", "notes": "No homes or other vulnerable buildings may stand where the yearly risk of death from a hazardous chemical site exceeds this. The EU's Seveso rules leave numbers like this to each country.", "evidence": "reported" }, { "name": "Rocket Launches", "kind": "Safety limit", "low": 0.0001, "high": null, "mid": 0.0001, "display": "1 in a million per launch", "reference": "US Code of Federal Regulations, 14 CFR 450.101", "url": "https://www.law.cornell.edu/cfr/text/14/450.101", "notes": "The most any one member of the public may be exposed to in a launch. Across everyone, expected casualties may not exceed 1 in 10,000 per launch.", "evidence": "published" }, { "name": "Mars Sample Return", "kind": "Mission requirement", "low": 0.0001, "high": null, "mid": 0.0001, "display": "1 in a million over 100 years", "reference": "ESA Earth Return Orbiter mission requirement, in ISSFD 2024 paper 10-6", "url": "https://issfd.org/ISSFD_2024/ISSFD2024_10-6.pdf", "notes": "The chance that a particle from Mars reaches Earth's biosphere, set because the harm, however unlikely, could reach everyone. It is the closest thing here to a limit on a planet-wide risk from something we choose to do.", "evidence": "published" }, { "name": "Human Driving (US)", "kind": "Observed rate, not a limit", "low": 1.2e-06, "high": null, "mid": 1.2e-06, "display": "About 1 death per 80 million miles", "reference": "NHTSA, early estimate of 2024 traffic fatalities", "url": "https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/813710", "notes": "1.20 deaths per 100 million miles driven. There is no legal limit; this is the record self-driving cars are usually measured against.", "evidence": "published" }, { "name": "Aircraft Systems", "kind": "Design target", "low": 1e-07, "high": null, "mid": 1e-07, "display": "1 in a billion per flight hour", "reference": "FAA Advisory Circular 25.1309-1", "notes": "A design target for a single catastrophic failure condition in a transport aircraft. It is a rate per hour flown, not per year.", "evidence": "reported" } ], "caveat": "These figures use different units and measure different things: a rate for one structure, a lifetime risk to one person, a chance per mission. p(doom) estimates a one-time civilizational probability. The comparison gives a sense of scale and is not an equivalence. We found no published numeric limit for medical devices or BSL-4 laboratories, so they are not shown.", "evidenceClasses": [ { "id": "published", "label": "Read in the published report", "desc": "We read the figure in the study or report itself." }, { "id": "own", "label": "Own words, read at the source", "desc": "We read the person's own post, essay or transcript." }, { "id": "reported", "label": "Reported by a named outlet", "desc": "A named outlet quotes or reports the figure; we did not reach the original statement." }, { "id": "cited", "label": "Cited through a compilation", "desc": "The figure reached us through Wikipedia or PauseAI with a stated source that we could not open or have not heard." } ], "forecastComparison": { "note": "Different panels and survey designs. Differences between rows are not a measured trend.", "rows": [ { "study": "XPT 2022", "outcome": "Human extinction", "superforecasters": 0.38, "experts": 3 }, { "study": "XPT 2022", "outcome": "Over 10% of humanity dies within five years", "superforecasters": 2.13, "experts": 12 }, { "study": "LEAP 2026", "outcome": "Over 10% of humanity dies within five years", "superforecasters": 2.4, "experts": 5 } ] } }