What is p(doom)?
The probability, in someone’s own estimate, that advanced AI ends in human extinction or a comparable catastrophe. Short answers below, each backed by the p(doom) database.
What does p(doom) mean?
p(doom) is shorthand for the probability that advanced AI leads to human extinction or a comparable, irreversible catastrophe. People write it like a probability, for example “my p(doom) is 10%”. It is a personal, subjective estimate, not a measurement.
Is p(doom) a real probability?
It is a degree of belief, not a frequency anyone can measure: there is no track record of AI catastrophes to count. It is still useful for comparing how worried different people are, and for noticing when someone changes their mind.
Why don’t p(doom) numbers agree?
Mostly because people answer different questions. They differ on scope (extinction, permanent disempowerment, or any catastrophe), on timeframe (this decade, by 2100, or ever) and on conditions (“if nothing changes”, “per generation of models”). Two people who say “10%” can mean very different things.
- Evan Hubinger: >10%, AI kills all humans, within the next decade
- Roman Yampolskiy: 99.999999%, AI catastrophe / extinction, no timeframe given
- Dario Amodei: 25%, things go really, really badly, no timeframe given
- Yann LeCun: ≈0%, existential risk from AI, no timeframe given
- Ryan Dahl: 0.01%, extinction, next 10 years
- Jaan Tallinn: 1–50% per 10× compute (~7% point estimate), life-ending disaster from AI, per generation of frontier training…, per 10× compute generation (he says…
What is the average p(doom)?
Across the 223 people in the database, the median of each person’s latest number is 10%. Expert surveys put it at 17.5% (median of the latest surveys), and the Manifold market “Will AI wipe out humanity before 2100?” sits at 13.8%. Combined, the p(doom) index is 13.0.
Who has the highest and lowest p(doom)?
The highest in the database is Roman Yampolskiy at 99.999999%. 37 people put it at or near zero, among them Abe Murray, Adam Holter, Ali Ghodsi, Andrew McAfee, Bill Karr, Chris Albon.
Who refused to give a p(doom)?
22 people were asked on the record and gave no number, including Donald Trump, Ted Cruz, Stuart Russell, Neel Nanda, Demis Hassabis, Sundar Pichai, Max Harms, Guillaume Verdon (Beff Jezos). Refusals are listed, never counted.
Does p(doom) depend on who is asked?
In this database, yes. Grouped by who the speaker works for, the median ranges from building AI 0.5%, investors 1.5%, independent 5%, government 7.5%, media 10%, ex-frontier labs 12.5%, frontier labs 12.5%, academics 15%, safety orgs 50%. That is a pattern in who says what, not proof of anyone’s motives.
What are the criticisms of p(doom)?
Critics say one number hides the scenario, the timeframe and the reasoning behind it; that it turns an open research question into a team badge; and that subjective numbers can’t be checked. Several researchers decline to give one for these reasons.
One well-known version: “Stop talking about p(doom)” (EA Forum).
How is the p(doom) database built?
Only numbers people stated themselves, in public, with a source a stranger can check. Ranges stay ranges, conditions are kept, a changed mind gets a new row, and refusals are listed separately. A red dot means a maintainer re-checked the quote against the primary source.
How do I add a p(doom) to the database?
Open an issue on the public GitHub repository with the person’s exact words and a link where anyone can confirm them. Submissions without a primary source are not added.
Is this connected to the $PDOOM coin?
Yes: the site is run by the p(doom) project, which also runs the $PDOOM memecoin on Solana. The data is independent open data; the coin changes no number and nothing here is financial advice.