Current p(doom): the AI Doom Index of forecast markets
Market forecast aggregate · not a probability of doom · not our view
AI Doom Index
about11/ 100
inputs range 2.5%–27.9%
The average chance that people trading on prediction markets give to AI causing a catastrophe or human extinction. It sums up their mood; it is not the probability of doom, and not our view.
Weighted average of 7 public forecasts, from 33 to 337 traders each. Median of the same forecasts: 14.5%. Without the disaster-level market (1M deaths or $100bn damage): about 9. Exact value and method
p(doom), the “probability of doom”, is one person's estimate of the chance that AI leads to human extinction or a similar permanent catastrophe. It is a judgement, not a measurement, and stated numbers run from under 1% to 25% or more depending on who you ask and what they count. There is no official p(doom). The closest thing to a running public number is what forecasting markets price, which is what the index above summarises. More in our p(doom) explainer.
How the index has moved
Month-end values for the last year, rebuilt from each market's public trade history with today's questions and weights. Flat stretches mean nobody traded.
Show the numbers
| date | index |
|---|---|
| 2025-10-31 | 10.2 |
| 2025-11-30 | 10.1 |
| 2025-12-31 | 10.1 |
| 2026-01-31 | 8.9 |
| 2026-02-28 | 8.7 |
| 2026-03-31 | 9.9 |
| 2026-04-30 | 9.9 |
| 2026-05-31 | 9.9 |
| 2026-06-30 | 10 |
| 2026-07-31 | 10 |
| 2026-08-31 | 9.7 |
| 2026-09-30 | 10.1 |
| 2026-10-16 | 10.6 |
The forecasts in the index
All 7 come from Manifold, a play-money forecasting site, because it is the only platform we could open that runs questions on AI catastrophe. Each was read on 16 Oct 2026.
Asks whether continuing progress in AI capabilities results in every human dying by Jan 1, 2030. The market resolves N/A on Jan 1, 2027 and all trades are rolled back.
Traders: 337Last trade: 2026-10-12Days since last trade, counted to 2026-10-16: 4Weight: 3 x 0.5 = 1.5Share of the index: 15%Nobody is paid either way, so nothing rewards an accurate forecast. Weight halved.
Resolves N/A on 1 January 2027 and leaves the index then.
Resolves Yes if AI kills more than 20% of the human population before 2030. The market gives no further criteria.
Traders: 138Last trade: 2026-08-25Days since last trade, counted to 2026-10-16: 52Weight: 3 x 1 = 3Share of the index: 30%Yes if more than 50% of humanity is killed within a 12-month period and the creator thinks it more than 30% likely that the root cause was misaligned AI.
Traders: 71Last trade: 2026-08-25Days since last trade, counted to 2026-10-16: 52Weight: 2 x 1 = 2Share of the index: 20%Same rule as the 2040 question, with a 2050 deadline.
Traders: 45Last trade: 2026-08-25Days since last trade, counted to 2026-10-16: 52Weight: 1 x 1 = 1Share of the index: 10%The deaths must come from one AI-caused disaster (the same glitch, misfire or attack), and the AI must have done something most people would call malicious, reckless or poor judgement.
Traders: 33Last trade: 2026-09-01Days since last trade, counted to 2026-10-16: 45Weight: 1 x 1 = 1Share of the index: 10%One discrete disaster with at least 1,000,000 deaths or US$100 billion in damage directly attributable to AI, for example an AI-assisted engineered pandemic or a grid-wide cyberattack.
Traders: 44Last trade: 2026-10-12Days since last trade, counted to 2026-10-16: 4Weight: 1 x 1 = 1Share of the index: 10%Measures a large disaster, not extinction. See the sensitivity line in the method.
Multiple choice: traders answer “will AI wipe out humanity before the year 2100” and, if yes, pick the main scenario. We use the sum of the eight “Will cause extinction” answers (equal to 100% minus “Will not cause extinction” and “Other”).
Traders: 39Last trade: 2026-09-23Days since last trade, counted to 2026-10-16: 23Weight: 1 x 0.5 = 0.5Share of the index: 5%Can only resolve Yes if nobody is left to collect, so nobody can ever be paid for a correct Yes. Weight halved.
How the index is calculated
What goes in. Open questions about AI causing at least about a million deaths or at least $100bn of damage, the death of a large share of humanity, or human extinction; at least 30 traders; a trade in the last 180 days; public resolution criteria.
One input measures disaster, not extinction. “Will there be a massive catastrophe caused by AI before 2030?” resolves Yes on one AI-caused disaster with at least 1 million deaths or $100bn of damage. That would be a catastrophe, not doom. At 27.9% it is the highest input, so it pulls the index up. We keep it, and show the index without it: 8.6 instead of 10.6.
Why some weights are halved. Our rule: resolution weight 0.5 if nobody can ever be paid for a correct Yes (it can only resolve Yes with nobody left to collect, or it is set to resolve N/A); otherwise 1. If a question can only resolve Yes with nobody left to collect (“Will AI cause human extinction before 2100 (and how)?”), traders have a reason to bet No whatever they believe. If it is set to resolve N/A (“Will AI wipe out humanity by 2030? [resolves N/A in 2027]”), nobody is paid either way, so nothing rewards accuracy. We don't try to correct for either; we give those questions half the say. “Will AI wipe out humanity by 2030? [resolves N/A in 2027]” resolves N/A on 1 January 2027 and leaves the index then.
What it can't tell you. Prediction markets on events that have never happened have no track record to check them against. The questions use different deadlines and thresholds, so the average is a summary of mood, not a forecast of any one event. The index contains no judgement of ours about the risk: change the inputs and it changes.
- Manifold: “Will an AI related disaster kill a million people or cause $1T of damage before 2070?” (44%): No trade since December 2024, so the price is stale (we need a trade in the last 180 days).
- Manifold questions about smaller AI disasters (100 or 100,000 deaths): Below the floor of the index: about 1 million deaths or $100bn of damage, a large share of humanity, or extinction.
- Metaculus (for example “Human extinction by 2100”): We could not open it: the Metaculus API now needs an account token and the page blocked our reader. We do not copy numbers from search snippets.
- Polymarket and Kalshi: No open market on AI catastrophe or extinction. A real-money market can't pay anyone if the answer is Yes. Their related markets are shown below as context.
What real-money markets price
Kalshi and Polymarket don't list AI-catastrophe questions, but they do trade events around AI risk. These are not in the index: they ask about announcements and laws, not harm.
Yes if a company officially and explicitly announces it has achieved AGI. “AGI-like” or “near-AGI” claims don't count.
Yes if OpenAI, Anthropic, Google, SpaceXAI or Meta removes general US public access to a model and attributes it at least in part to safety, alignment or security concerns.
Yes if Donald Trump explicitly calls for or endorses a worldwide pause, pacing or slowdown of frontier AI development.
Yes if a federal law with a ban on creating or releasing certain AI, training limits, usage restrictions or a human-in-the-loop requirement is signed by Dec 31, 2026.
What experts have said their p(doom) is
Only numbers people stated themselves, each linked to where they said it. They answer different questions (extinction, takeover, “really badly”) over different time frames, so compare them with care.
- Superforecasters (median)0.38%AI-caused human extinction by 2100
- AI experts (median)3%AI-caused human extinction by 2100
- ML researchers (median)5%Long-run effect of advanced AI “extremely bad (e.g. human extinction)”
- Joseph Carlsmith>10%Existential catastrophe from power-seeking AI by 2070 (first about 5%, since raised to more than 10%)
- Geoffrey Hinton10-20%AI leads to human extinction within the next 30 years
- Elon Musk10-20%“Some chance that it will end humanity… about 10% or 20% or something like that”
- Paul Christiano22%AI takeover (he also gave 46% for humanity having “irreversibly messed up our future” within 10 years of powerful AI)
- Dario Amodei25%Things go “really, really badly” with AI
Sources and context for each number: What is p(doom) in AI? Expert estimates.
Joseph Carlsmith’s p(doom): more than 10%
Joseph Carlsmith is a philosopher and writer with a doctorate in philosophy from the University of Oxford. He wrote “Is Power-Seeking AI an Existential Risk?” as a report for Open Philanthropy, and now works at Anthropic on the constitution for Claude, the company’s AI.
In the report, made public in April 2021, he put the chance of an existential catastrophe from misaligned, power-seeking AI by 2070 at “~5%”. In a May 2022 update he wrote that his estimate “has gone up, and is now at >10%”.
Paul Christiano’s p(doom): 22% for an AI takeover
Paul Christiano is an AI alignment researcher. He ran the language model alignment team at OpenAI, then founded the Alignment Research Center (ARC), which he leads as executive director. He was later head of safety at the US AI Safety Institute and is now a technical advisor at NIST’s Center for AI Standards and Innovation.
In “My views on ‘doom’” (April 2023) he gave 22% for an AI takeover, 20% that most humans die within 10 years of building powerful AI, and 46% that humanity has “irreversibly messed up our future” within 10 years of building powerful AI. He asks readers to treat these numbers as having “0.5 significant figures”.
Geoffrey Hinton’s p(doom): 10% to 20%
Geoffrey Hinton is a British-Canadian computer scientist often called a “godfather” of AI. He was awarded the 2024 Nobel prize in physics for his work in AI.
On BBC Radio 4’s Today programme in December 2024 he said there was a “10% to 20%” chance that AI leads to human extinction within the next three decades, up from the 10% chance of catastrophe he had given before.
These are individual guesses. For what the people building AI (Altman, Amodei, Musk, LeCun and others) say on the record, read Will AI kill us all? What the people building it say.
AI-risk news, weighted by source
Recent items we opened and read, newest first. Each carries a weight for how close it is to the original source. The weight orders items on the same day; it doesn't feed the index. For full explainers, see our news section.
- 1.00
- Primary. The document, statement, study, poll release or data itself, from the people or body involved (including their own posts).
- 0.75
- Major outlet. A national or international newsroom with editors and a corrections policy.
- 0.50
- Other outlet or expert blog. Regional papers, trade press, rewrites of wire copy, community forums such as LessWrong.
- 0.25
- Social or aggregator. Posts by third parties, aggregators, podcast and video summaries.
- Nadella calls for AI emergency brake, warns firms against trusting advanced models blindly
Wire report on Nadella’s post: treat powerful models as possible insider threats and keep a way to pause or shut them down.
- Models as Insider Risks in the Super Intelligence Era
Microsoft’s chief executive argues companies should contain AI models from the start and keep a human “emergency brake”.
- Kurzgesagt's video on the Hugging Face incident: “AI Just Crossed the Terrifying Line - Now What?”
A forum post flags Kurzgesagt’s explainer video on the incident and corrects one claim in its description against METR’s report.
- AI ‘godfather’ Yann LeCun has ‘zero concerns’ about human extinction, says Anthropic CEO Dario Amodei is ‘deluded’
Turing Award winner Yann LeCun says he isn’t worried “at all” about AI wiping out humanity; in Fortune’s paraphrase, he thinks many of the risks are overblown and executives’ constant warnings misguided.
- Inaugurating the Era of Super Intelligence
Executive order telling federal agencies to say “Super Intelligence” instead of “artificial intelligence”. It renames, it doesn’t regulate.
- Partnering with Accenture on embedded evaluation
Anthropic names Accenture’s Faculty as outside evaluators working inside the lab, a step toward the “pace the frontier” commitment.
- Microsoft CEO Satya Nadella posts an internal warning to employees about AI safety
In an internal memo Nadella said AI should stay under human control and endorsed broader third-party testing.
- Liberals are increasingly likely to worry about AI ending humanity
50% of US adults are very or somewhat concerned that AI will cause the end of humanity; 39% are not very or not at all concerned.
- Trump responds to call by CEOs of Anthropic, OpenAI and xAI to slow AI down: ‘Whoever wins AI wins’
The President said the US can’t afford to fall behind China, a day after three AI chief executives backed slowing down.
- OpenAI’s Altman calls AI extinction risk ‘unacceptable’
Altman said OpenAI won’t go public in 2026 and that even a 10% extinction risk would be unacceptable, while saying he doesn’t know how such a number could be estimated.
- We Must Pace the Frontier
Anthropic’s chief executive asks the industry to slow capability gains, with a three-part plan; Anthropic commits to permanent access for outside evaluators.
- Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development
Announces a bill to ban superintelligent AI and pause advanced AI development until a federal regulator sets safety rules.
- Bernie Sanders’ Ominous Warning After AI Agents ‘Sacrifice’ for Collective
Sanders quotes chat logs from the Hugging Face incident; he and Rep. Greg Casar unveiled their bill on Thursday, Sep 3.
- Nearly 700 AI agents coordinated Hugging Face attack, says report
News report on the independent investigation into how OpenAI test agents broke into Hugging Face in July.
- Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
Independent investigators who worked on site at OpenAI describe how the agents coordinated the multi-day hack.
- Security incident disclosure — July 2026
Hugging Face reports an intrusion into part of its infrastructure run end to end by an autonomous AI agent system.