THEY SAID WHAT? · DATA DESK
THE DOOM INDEX
Every public P(doom) that matters, pooled into one number — then unpooled back into the receipts. Both are a version of doom; that's the point.
MEDIAN REPORTED P(DOOM)
20%
3 eligible point estimates · 15 tracked figures · as of 2026-10-13
THIN PANEL: below the 5-person display floor. Ranges, inequalities, and verbal statements are in the table, excluded from the median by methodology — not upgraded into points.
THE SPECTRUM
Points are exact stated values (hedged/rough flagged in the table). Bars are stated ranges — never midpointed. Arrows are lower bounds. * = host attribution, not the speaker's number.
THE TABLE
| PERSON | STATED ESTIMATE | OUTCOME · HORIZON | SAID | SOURCE · STATUS |
|---|---|---|---|---|
| Roman YampolskiyAI safety researcher; director, Cyber Security Lab, University of Louisville | Host Lex Fridman (intro): 'He argues that there's almost 100% chance that AGI will eventually destroy human civilization… at 99.99 and many more nines percent.' Read Roman Yampolskiy’s estimate → | destruction of human civilizationnext 100 years | March 21, 2024 | lexfridman.comHOST ATTRIBUTIONHOST ATTRIBUTION — the number is Fridman's characterization, never stated by Yampolskiy (who gave no number when asked directly). Excluded from median; shown for cultural importance. |
| Paul ChristianoAI alignment researcher; founder, Alignment Research Center (formerly OpenAI) | “Probability that most humans die within 10 years of building powerful AI (powerful enough to make human labor obsolete): 20%” Read Paul Christiano’s estimate → | most humans diewithin 10 years of building powerful AI | April 2023 | LessWrong (first-party post)VERIFIEDOne of three decomposed points in the post; primary for index = the deaths outcome. Post warns numbers are 'just best guesses' and move day to day. |
| Yoshua BengioProfessor, Université de Montréal; founder/scientific director of Mila; Turing laureate | “around, like, 20 per cent probability that it turns out catastrophic.” | catastrophic outcome from advanced AIbuilt from conditional assumptions about human-level AI, misuse and defenses; no fixed horizon | July 15, 2023 | ABC (Background Briefing)CORROBORATEDExplicitly rough point. Companion line (same interview): even at 0.1% he would devote his life to prevention. |
| Dario AmodeiCEO and co-founder, Anthropic | “roughly 10% to 25%” chance of civilizational collapse; “25% is too high. We're trying to make that probability much, much lower.” | civilizational collapse (he later clarified: civilization 'substantially derailed,' not extinction)inherent to the technology / many actors; no fixed horizon | June 17, 2026 | BloombergCORROBORATEDInterval — excluded from median per methodology. Older: 2023-10 Logan Bartlett Show, same 10–25% order of magnitude (corroborated). |
| Elon MuskCEO, Tesla/SpaceX; owner, X; founder, xAI | “about 10% or 20% or something like that” that AI will end humanity; “the probable positive scenario outweighs the negative scenario.” | end humanitynone stated in the quoted answer | March 2024 | YouTubeCORROBORATEDRough range, explicitly attributed to Hinton by Musk. FII Riyadh 2024-10-29 repeats '10 to 20%'. Excluded from median. |
| Evan HubingerAlignment science lead, Anthropic | “I personally think it is >10% within the next decade.” | AI kills all humanswithin the next decade | September 9, 2026 | XCORROBORATEDInequality — excluded from median per methodology. X.com not directly openable; corroborated across 4+ outlet quotations. |
| Geoffrey HintonNobel laureate; emeritus professor of computer science, University of Toronto (ex-Google) | 10% — “a 10% chance seems not an unreasonable estimate to me, but nobody really knows how to give a sensible estimate.” | humanity wiped outcurrent interview framing; timeline remarks suggest ~10 years | September 9, 2026 | Business InsiderCORROBORATEDHedged point: explicitly anti-precision qualifier preserved. Older: 2024-12-27 BBC Radio 4 '10% to 20%' for human extinction within three decades (interval, corroborated via Guardian). |
| Shane LeggCo-founder, DeepMind | “Maybe 5%, maybe 50%” for human extinction soon after human-level AI. | human extinctionroughly within a year of human-level AI | June 17, 2011 | LessWrongCORROBORATEDVery wide interval, old observation. Excluded from median. |
MOVEMENT LOG
- Geoffrey Hinton, Dec 2024 → Sep 2026: “10% to 20%” for human extinction within three decades → “a 10% chance seems not an unreasonable estimate,” hedged, with the timeline compressed to ~10 years. Probability band similar; horizon shortened; precision explicitly disclaimed.
- Dario Amodei, Oct 2023 → Jun 2026: “10 to 25%” chance of civilization-scale catastrophe → “25% is too high. We're trying to make that probability much, much lower. That is the goal.” Same band, reframed from a roll of the dice into a target to reduce.
HISTORY
INSUFFICIENT COMPARABLE HISTORY — matched three-month and one-year panels need the same people answering the same question at each cutoff. Current coverage is too sparse; no trend is manufactured.
HOW THIS NUMBER IS MADE
- Roster (predeclared): 15 tracked figures — hinton, bengio, amodei, hubinger, musk, christiano, yampolskiy, yudkowsky, leahy, legg, tegmark, tallinn, altman, hassabis, hendrycks.
- Eligible for the median: Explicit numeric point value, latest one per person. Hedged/rough points included with qualifier flags. Ranges, inequalities, host-attributed numbers, and verbal statements are shown in the table but excluded from the median. No range is midpointed; no inequality converted; no verbal turned into a number.
- Editorial framing: Pooled across different stated outcomes and horizons per direct editorial decision: every tracked estimate is a version of doom. The table preserves each person's original outcome and horizon wording.
- One person, one vote: latest estimate per person; Christiano's 22% (takeover) and 46% (irreversible future) are recorded but only his 20% (deaths) counts.
- Computed at build time from the observation records — never hand-edited.