Methodology

How the index works

Who is listed

We list a person when all three of these are true. We apply them the same way to everyone, whatever their view, so the list doesn’t depend on whom we happen to like.

  1. An explicit number. They have publicly stated a number or range for the probability that AI leads to human extinction or a permanent, catastrophic loss of human control. We never infer a number from tone. When someone’s number is about something narrower or broader (a time limit, “things going really badly”), we note it next to their entry.
  2. A source. Every estimate links to where it was said. We prefer the original interview, podcast, post, talk or paper; when we could only find a news report, the link says “reported”.
  3. A qualifying role in one of the five groups below, now or in the past.

People with a qualifying role who haven’t stated a number are on our watchlist.

Each name links to the person’s Wikipedia article, for a neutral, independently edited bio. Where there is no article, it links to the profile the person points to themselves.

The five groups

Each person is in exactly one group: the first one in this order that their role fits.

  1. AI industry leaders: CEO, president, co-founder or chief scientist of a company whose main product is AI models.
  2. Government & policy: senior public office (minister, legislator, head of an agency or regulator), or head of an AI policy organisation.
  3. AI safety researchers: people whose main work is AI safety or alignment, at a safety organisation, on a lab’s safety team, or in academia.
  4. AI & computer science researchers: faculty or senior researchers in AI, computer science or cognitive science with a Turing Award, a Nobel Prize, or more than 10,000 citations.
  5. Public voices: writers, podcasters, investors and technologists outside the groups above, with an English Wikipedia article about them or an audience of more than 100,000.

The index

We take each person’s most recent estimate. A range counts at its midpoint (10–20% counts as 15%), and “more than X%” counts as X%, since that is all the person said. When one source gives several numbers, we use the one closest to our definition and keep the others on record. Estimates more than five years old are shown but not counted.

We take the median within each group, then the median of those group medians. Each group has the same weight however many people are in it, so a large group with similar views can’t outvote the rest, and adding more people of one kind moves only their group. A group needs at least three counted members to be part of the index. We use medians rather than averages so that a few very high or very low estimates don’t dominate.

The index is recomputed whenever the list changes. Any change to these rules will be dated and explained here.

Markets

Forecasting platforms and prediction markets show what people betting or forecasting on these questions think. Their numbers are shown for comparison and are not part of the index. Long-range questions such as “before 2100” are hard to bet on fairly, because a market can’t pay out if everyone is gone; read them as a rough signal.

Reuse

The index and data are free to reuse under CC BY 4.0: credit p(doom) Tracker with a link. There’s a free JSON endpoint and an embeddable badge.

Corrections

If we’ve misquoted someone or missed an estimate, email questions@pdoomtracker.com or suggest an estimate, and we’ll fix it. Corrections are recorded in the change log.