Metaculus AI Existential Risk Explorer
What the Metaculus crowd expects from AI and who that crowd is
Metaculus has asked about AI catastrophe and human extinction since 2017. This page shows how those forecasts have moved, how the community prediction compares with Metaculus Pro Forecasters and medalists, and how much the same people are behind each question.
Every line is the community prediction as Metaculus computes it: each person's current forecast, weighted toward the most recent. Group lines use the same rule for just that group.
Forecasts over time
What stands out
Figures with a dotted underline are computed from the latest data and change as forecasts do.
Who forecasts these questions
Everyone with a standing forecast. Pros are Metaculus's paid professional forecasters; medalists have won a Metaculus leaderboard medal (Pros are counted with Pros).
Comparing each group's current forecasts
Current values by group. The stretched axis spreads out both ends, so 1% and 5% sit as far apart as 50% and 84%.
Do the forecasts hang together?
Some of these questions are more stringent versions of other questions, meaning the probabilities should be lower on the more stringent version. This section checks whether the community prediction or subsets of forecasters who predicted both questions are logically consistent.
Uncertainty and disagreement
Recency-weighted forecasts for each group, showing the 25th-percentile forecaster's prediction, the median (that group's community prediction) and the 75th percentile. Disagreement is how far the typical forecaster in the group sits from the group's community prediction: the recency-weighted median earth mover's distance between each forecast and the aggregate. On a binary question that reflects the simple difference between the forecasts in percentage points. For the five-worlds question it is the share of probability you would have to move.
Picture each forecaster's beliefs as a mound of dirt. On a yes/no question there are only two places to pile it, Yes and No, so a 30% forecast puts 30% of the dirt on Yes. On the five-worlds question there are five piles, one per world. The group's community prediction is another set of piles. Earth mover's distance asks how much of the dirt you would have to move to turn a typical forecaster's piles into the group's. On a yes/no question that is just the gap between the two probabilities: a forecaster at 10% is 20 points from a group at 30%. On the five-worlds question it is the share of probability that has to change worlds: a forecaster at 50 / 30 / 20 against a group at 30 / 40 / 30 has to move 20 points. Low means most forecasters' piles look like the group's: broad agreement. High means people have piled their dirt in very different places: real disagreement.
Discover more
This analysis focuses on a subset of the Metaculus extinction questions with the most forecasters and overlap, but others exist on the platform as well and can be found collected in the AI Existential Risk question series.
You can also explore our long-running Ragnarök Series, which explores the risks from different sources. You can also find a nice visualization of the Ragnarök Series at the Possible Worlds Tree.
Methodology
Who counts. Everyone whose latest forecast on a question is still standing, meaning not withdrawn or expired. Bots, spam-flagged accounts and accounts excluded from aggregation are left out, as on Metaculus.
Community prediction. Metaculus's recency-weighted median, reproduced from the raw forecasts; it matches every community prediction Metaculus has stored for these questions. A group's line applies the same weights to that group's members only, the way Metaculus's aggregation explorer does, so in a small group the most recent forecasters carry most of the weight.
Groups. Pros are people who have been Metaculus Pro Forecasters. Medalists are people who have won a Metaculus leaderboard medal and are not Pros; Pros are shown as their own group everywhere on this page. Full-set forecasters are the people with a standing forecast on every question on this page.
Inferred chart. The implied chance that AI causes extinction multiplies each group's three Ragnarök forecasts at each moment. Nobody forecast that combined number directly.
“Extinction” is shorthand in some titles: the 2100 question means no humans alive on 1 January 2100; the Ragnarök question means a population decline of 95% or more.
Data: Metaculus, read from a replica of the live database. Current as of .