p(doom) is a person’s subjective probability that advanced artificial intelligence leads to “doom”: human extinction, or a permanent and severe loss of humanity’s ability to determine its own future. It is written as a number between 0 and 1, or as a percentage, and it expresses a degree of belief, not a measured frequency. The AIciety p(doom) Clock collects such estimates from its readers and shows their average.
Where the term comes from
The notation borrows from probability theory, where p(x) denotes the probability of an event x. “Doom” entered AI discourse through the existential-risk debates of the 2000s and 2010s, in which researchers such as Nick Bostrom and Eliezer Yudkowsky argued that a sufficiently capable AI with goals misaligned from human values could cause irreversible catastrophe. The shorthand p(doom) spread through the rationalist and AI safety communities and became mainstream vocabulary in 2023, when the release of ChatGPT turned the question of AI risk into a public one and journalists began asking researchers and executives for their number.
How it is used
Asking for someone’s p(doom) is a way of compressing a worldview into a single figure. The number reveals which arguments a person weighs, which evidence they find credible and how they handle deep uncertainty. Stated values among researchers range from well under one percent to above fifty. The 2023 Expert Survey on Progress in AI by AI Impacts, with 2,778 respondents, found a median estimate of 5 percent for human extinction or similarly permanent and severe disempowerment – unchanged from the previous year – and a median of 10 percent for humanity losing the ability to control future advanced AI systems.
The verb that often accompanies the term, to “up” or “update” one’s p(doom), reflects the Bayesian idea that a probability should move with new evidence. Publicly revising the number is a form of intellectual accountability: it can be held against you later.
What it is not
p(doom) is not a frequency: there is no series of worlds in which one could count how often things went wrong. It is not a measurement: no instrument reads it off, and the International AI Safety Report 2026, the most comprehensive scientific assessment of AI risk, deliberately states no number. And it is not a countdown: it describes no date, only a degree of belief.
Contested points
The term hides three ambiguities. First, what counts as “doom” varies between speakers: some mean extinction only, others include permanent disempowerment or any civilisation-scale catastrophe. Second, the time horizon is often left unstated; a lifetime estimate and a ten-year estimate are different claims. Third, the number may be conditional on current trajectories or unconditional on everything, including future interventions. Comparing two people’s p(doom) without settling these points compares different questions.
This site’s p(doom) Clock fixes all three: doom means extinction or permanent, severe disempowerment caused by AI; the horizon is the end of 2042; and the estimate is unconditional. It shows the arithmetic mean of a self-selected community’s estimates – an opinion, not a measured risk.
Related
FOOM · Paperclip Maximizer · Sharp Left Turn · Roko’s Basilisk · Chinese Room · I’m Upping My P(doom) — when the machine sings of its own undoing (essay on the song that made the term sing) · A Field Guide to I’m Upping My P(doom) · Nothing Went Foom! — when the machine argues for itself (the answer song) · p(doom) Clock · AI Pause · Accelerationism
Sources: IEEE Spectrum on the 2023 AI Impacts survey · AI Impacts: How bad a future do ML researchers expect? · International AI Safety Report 2026
Leave a Reply