p(doom)

The notation was circulating in specialist forums by March 2022. The earliest exact-dated use opened for this review is Vivek Hebbar's 26 March Alignment Forum question about “inside views” versus “betting odds”; this is an evidence boundary, not a coinage claim. A July MIRI newsletter associated the phrase with Yudkowsky's high concern, while ABC introduced it to a broad audience in July 2023. Later policy and economics publications show continuing use. No reviewed source establishes Yudkowsky as sole author.

Sources: s1, s2, s3, s4, s5, s8

The shorthand can reveal someone's rough level of concern, yet comparisons invite false precision when the proposition is missing. One speaker may mean extinction after superintelligence under present policy; another may include permanent disempowerment, misuse, or any future catastrophe. Identical numbers can therefore encode different causal models. Decision-relevant use should state the outcome, horizon, conditions, intervention assumptions, evidence, uncertainty range, and what would update the estimate. CSET argues that deep ignorance can make a lone probability inadequate for risk analysis.

Sources: s4, s8

“My p(doom) is 10%” is incomplete. A clearer statement might estimate “the chance that AI advances cause human extinction or similarly permanent, severe disempowerment within the next 100 years,” then describe assumptions and uncertainty. The 2023 survey of 2,778 AI authors separated differently worded questions and reported framing effects, illustrating why casual values are not automatically comparable. A prediction-market price is different again: CFTC describes it as an aggregate trading signal for a stated event contract, with terms and resolution, not one person's private credence.

Sources: s6, s7

Do not average or rank named people's p(doom) values unless their outcomes, horizons, conditions, and elicitation methods match. Interview percentages are not automatically forecasts generated by a model, and one unresolved existential event cannot supply a routine calibration score. The number also says little about causes or remedies. Use explicit scenario probabilities, decomposed pathways, sensitivity analysis, or resolvable forecasts when a decision requires more than a conversational shorthand.

Sources: s4, s6, s8