Numbers people actually said
A register of explicit probability estimates for catastrophic or existential outcomes from advanced AI. Each entry keeps the person, the number, the date, the exact question and the venue.
- These numbers are not comparable to each other. They answer different questions over different horizons: extinction within 30 years is not the same question as any catastrophic outcome this century.
- A P(doom) figure is a statement of belief, not a measurement. Nobody has a frequency table for this.
- Every row is flagged primary, reported or unverified. Seed entries were human-entered from widely reported statements and carry reported until a primary source is attached. Nothing here should be quoted as a primary citation while it says otherwise.
- Inferred positions live on the People pages and never appear on this chart.
One point per person: their most recent estimate on record. Bars show a stated range where one was given. 1 person is on record in the table below but not plotted, because the entry is flagged unverified and plotting it would present a placeholder as a stated number.
Every estimate on record
| Person | Probability | Question | Date | Source | Flag |
|---|---|---|---|---|---|
| Eliezer Yudkowsky Machine Intelligence Research Institute |
95% 90–99% |
Probability that the development of superintelligence under current conditions kills everyone He has generally declined to give a precise number, describing the outcome as near-certain absent a halt. The figure here is an interval reconstructed from his own phrasing and should be read as such. |
2023-03-29 | TIME magazine guest essay, 29 March 2023, and repeated public statements | reported |
| Yoshua Bengio Mila, Université de Montréal |
20% | Probability of catastrophic outcomes from advanced AI Bengio has described his own estimate as around 20 percent while stressing the wide uncertainty and arguing the number matters less than the fact that it is not negligible. |
2024-01-01 | Public interviews and written testimony, 2023 to 2025 | reported |
| Dario Amodei Anthropic |
20% 10–25% |
Probability that things go really badly with advanced AI Stated as a range rather than a point estimate, and about civilisational-scale bad outcomes rather than extinction specifically. |
2023-08-08 | Dwarkesh Podcast interview, August 2023 | reported |
| Geoffrey Hinton University of Toronto (formerly Google) |
15% 10–20% |
Probability that AI causes human extinction within the next 30 years Widely reported as a revision upward from the 10 percent figure he had used previously. |
2024-12-27 | BBC Radio 4, Today programme (guest editor interview), 27 December 2024 | reported |
| Elon Musk Tesla, SpaceX, xAI |
15% 10–20% |
Probability that AI annihilates humanity Stated alongside the claim that the most likely outcome is nonetheless positive. |
2024-03-18 | Abundance Summit, Great AI Debate, March 2024 | reported |
| Geoffrey Hinton University of Toronto (formerly Google) |
10% | Probability that AI wipes out humanity in the next 30 years The earlier figure against which the 2024 statement was reported as a change. |
2023-05-01 | Interviews given around his departure from Google, May 2023 | reported |
| Toby Ord University of Oxford |
10% | Probability of an existential catastrophe from unaligned artificial intelligence this century Given as roughly 1 in 10 for AI specifically, within a total existential risk estimate of about 1 in 6 for the century. |
2020-03-24 | The Precipice: Existential Risk and the Future of Humanity, Bloomsbury, 2020 | reported |
| Yann LeCun Meta |
0% | Probability that AI causes human extinction LeCun has not given a formal number. He has repeatedly described the risk as negligible and the framing as unscientific. The value here is a placeholder for 'asserted negligible' and is flagged accordingly rather than presented as his figure. |
2023-10-01 | Repeated public statements on social media and in interviews, 2023 to 2025 | unverified |
A numeric probability, given by a named person, about a catastrophic or existential outcome from AI, with a date and an identifiable venue. Ranges are stored as a low and a high with the midpoint used for plotting. A qualitative statement ("very likely", "negligible") is not an entry, however strongly it is worded. Where a person has repeatedly declined to give a number, that absence is itself recorded on their page rather than filled in.