p(doom)

"AI risk" names a whole field of concerns: bias, misuse, labour displacement, concentration of power. A p(doom) fixes on one outcome, catastrophe for humanity, and it stays attached to the person who stated it. The second cut matters more on a first reading. A p(doom) is a credence, not a forecast with a track record and not a measured frequency. Nobody took a reading. Two stated numbers are not comparable unless both speakers fixed the same definition of doom and the same time horizon, and usually they have not.

The term comes out of the rationalist community and LessWrong, and it reached mainstream coverage in 2023. It has no documented coiner and no origin date, which is unusual in this vocabulary and tells you what kind of word it is: forum and lab-corridor shorthand that spread, not a published metric with a method behind it. Figures attach to names instead of procedures. Dario Amodei has put his own at around 25%, reported by Axios on 19 September 2025.

What the number is for

The twins table in The p(doom) song, annotated pairs p(doom) with a 1942 case. At a Berkeley meeting in July 1942, Edward Teller raised the possibility that a fission weapon might ignite the atmosphere. Oppenheimer took the worry to Arthur Compton in Chicago, and Compton decided he would stop the project if the chance of vaporising the Earth came out above roughly three in a million. The calculations came out below that threshold and the work continued.

The figure carried a decision rule, so it governed action instead of describing a mood. It is also a recollection: the three-in-a-million value comes from Compton's 1959 account to Pearl Buck, seventeen years after the meeting, not from a document written at the time. Compton's own 1956 memoir covers the episode without the number.

So two questions are worth asking of any stated p(doom). What definition of doom and what time horizon does it cover, and what would the speaker do differently at half the number.