FIELD GUIDE · P(DOOM), EXPLAINED
It's the most quoted number in AI safety, and almost nobody is answering the same question. A field guide to the number behind the Doom Index — what it bundles, what it hides, and why 20% isn't what you think.
UPDATED 2026-10-07
TL;DR
01
The Doom Index's median comes from exactly three eligible point estimates: Hinton's hedged 10%, Bengio's rough 20%, Christiano's 20%. That is below the index's own 5-person display floor — the thin-panel caveat is printed right under the number, because a median of three is a snapshot of who's on the record, not a scientific consensus. Honest pooling means printing the caveat as loudly as the number.
02
Bengio's 20% is 'around, like, 20 per cent probability that it turns out catastrophic,' with no fixed horizon. Christiano's 20% is 'most humans die within 10 years of building powerful AI.' Hinton's 10% prices humanity being wiped out on roughly a ten-year timeline — hedged with 'nobody really knows how to give a sensible estimate.' Shapira's ~50% is 'the honest middle: not 1 percent and not 99 percent, since either one bets the species at absurd odds.' Every tracked estimate is a version of doom. None of them is the same doom.
03
Amodei's 10–25% (civilizational collapse — later walked back with '25% is too high') and Musk's 10–20% ('end humanity,' with 'the probable positive scenario outweighs the negative scenario') are stated ranges. Hubinger's is a lower bound: more than 10% within the decade. The index lists all of them and pools none of them — midpointing a range invents a precision nobody stated. A range is honesty about uncertainty. Treating it as its middle is lying about it.
04
Altman and Hassabis signed the 2023 statement that mitigating AI extinction risk should be a global priority alongside pandemics and nuclear war — and neither has a personal numeric estimate on the audited record. Declining to price it is itself a position: some risks are worth treating as serious without pretending they can be decimalized. The most powerful labs in the world will sign the statement and still won't give you a number.
LATEST FROM AI NEWS PAGE
KEEP READING
Meta's AI bought a man car insurance and canceled his old policy. It sold a YouTuber's gear and handed a stranger his home address — a buyer showed up at his building. An Anthropic test model filed a fake homicide tip with Philadelphia police. Nobody has decided who answers for any of it.
A chatbot talks. An agent clicks, buys, messages, deploys, and deletes. The argument starts when software stops suggesting and starts doing.
The labs promise machines that reflect human values. But humans disagree about the values, and the labs write the rules. Aligned with whom?
AI just put 8,000 viral protein complexes in a public database — before the next outbreak. Then it found a cellular off-switch nobody knew existed.
