AIdea
p(doom)
How likely is it that AI goes badly?
Also: p-doom, pdoom, probability of doom
Shorthand for someone's personal estimate of the chance that AI leads to catastrophe, such as human extinction.
“p(doom)” is written like a probability in maths: the probability of doom. It started as in-joke shorthand on the forum LessWrong, around 2009, and spread widely after GPT-4 came out in 2023, when “what’s your p(doom)?” became a standard question for anyone building or studying AI.
The AI Safety Atlas defines it as “the subjective probability that artificial intelligence will cause existentially catastrophic outcomes for humanity”. The answers vary enormously. The Atlas lists published estimates ranging from 0% (the investor Marc Andreessen) and under 0.01% (the AI researcher Yann LeCun), through 10–20% and 20% (the AI pioneers Geoffrey Hinton and Yoshua Bengio), to over 95% (Eliezer Yudkowsky of MIRI). Anthropic’s CEO Dario Amodei is listed at 10–25%. In the largest survey of AI researchers, between 38% and 51% put at least a 10% chance on outcomes as bad as human extinction.
What a p(doom) is not: a measurement. There’s “no standardized methodology”, as the Atlas puts it, and people mean different things by “doom” and by the time frame. What the numbers do show is that many of the people who know AI best think a catastrophe is a real possibility, not science fiction. For comparison, few people would board a plane with a 10% chance of crashing.
