The best-documented comparison of expert views on whether AI could end humanity comes from the Existential Risk Persuasion Tournament, run from June through October 2022 by the Forecasting Research Institute with 169 forecasters. The median superforecaster put the chance of AI-caused human extinction by 2100 at 0.38 percent, while the median domain expert put it at three percent, a gap of about one order of magnitude. Both numbers are informed opinions about a future event, and months of structured debate did not bring the two groups together.
Key takeaways
- The tournament compared 89 superforecasters with 80 domain experts, and AI was the topic on which their estimates differed most.
- For an AI-caused catastrophe killing at least 10 percent of humans within five years, the medians were 2.13 percent (superforecasters) and 12 percent (domain experts).
- Neither group moved toward the other after reading each other's arguments, and a later re-survey of skeptics found no noticeable increase in their concern.
- In a separate public survey, the answer changed drastically depending on how the question was asked, which shows how fragile such numbers are.
- These are forecasts, not measurements. No one can check an estimate about 2100 until 2100.
How the tournament worked
The Forecasting Research Institute (FRI) designed the tournament to reward forecasters for giving accurate estimates, writing explanations, persuading one another, and updating their views. The authors, among them Ezra Karger and Philip Tetlock, published a working paper that was first released on July 10, 2023 and revised on August 8, 2023. Nearly 200 people signed up for the four-month exercise, 169 submitted forecasts, and 111 took part in all four stages.
The two main groups were chosen for different reasons. Superforecasters are people who scored highly in earlier forecasting tournaments on short-run questions. The 80 experts were recruited from organizations and academic departments that work on existential risk, which includes AI. The design lets the paper ask a pointed question: when people with a strong record on near-term predictions and people with subject knowledge disagree, which group should be given more weight?
The numbers
The table gives the final median forecasts reported in the working paper. Extinction in the AI questions meant humanity going extinct or the world population falling below 5,000. A catastrophe meant deaths of more than 10 percent of humans alive at the start of a five-year period. Square brackets show the paper's 95 percent confidence intervals.
| Question (by 2100) | Superforecasters | Domain experts |
|---|---|---|
| AI-caused extinction | 0.38 percent [0.10, 0.75] | 3 percent [0.49, 10.00] |
| AI-caused catastrophe | 2.13 percent [1.83, 3.00] | 12 percent [4.0, 18.5] |
A public survey of 912 college graduates, run in parallel, produced a median of two percent for AI-caused extinction, between the two expert groups. The paper also reports that both groups saw AI as the most likely cause of extinction among the options in the tournament, even though they disagreed on the size of the risk.
Why the groups did not converge
The authors treat the lack of convergence as the central puzzle. Forecasters had months, incentives to persuade each other, and team discussions, yet the two groups stayed apart. The paper links part of the split to timelines: the median expert expected the first publicly known "advanced AI" system by 2046 and the median superforecaster by 2060, using a definition that combines a two-hour adversarial Turing test with benchmark performance. Superforecasters were also more doubtful about very large economic effects of AI.
The paper reports a pattern across participants, too. Within both groups, those who were best at predicting what other participants would say tended to give lower estimates of catastrophic and extinction risk. The authors split participants into thirds by their AI extinction forecast: the most concerned third had 17 superforecasters and 39 experts, and the least concerned third had 39 superforecasters and 17 experts. After new language models appeared, FRI re-surveyed some of its most skeptical participants and saw no evidence that they had become noticeably more concerned.
What the gap does not show
The numbers cannot settle who is right. A forecast about 2100 has no feedback until then, so a track record on short questions only suggests, and does not prove, accuracy on a century-long one. The tournament ended in October 2022, so the forecasts predate systems released afterward. The authors also describe it as the first iteration of the exercise.
Elicitation matters as well. In a subset of 405 public respondents who answered with two methods, a plain text box gave a median of 1.5 percent for AI extinction, while an alternative method gave 1 in 40 million. The same people, on the same topic, produced answers separated by several orders of magnitude. That is a reason to read any single risk percentage, including the ones above, as a rough statement of concern and not as a measured probability.
Questions to ask of any AI risk number
- Who produced the estimate, and how were they chosen?
- What exactly counts as the event: extinction, a lasting loss of control, or a catastrophe with a stated death toll?
- By what year, and from which starting point?
- Was it a median, a mean, or a single person's view, and was a range given?
- How was the question asked, and would a different wording change the answer?
- Can the claim ever be checked, or does it depend only on argument?
This article gives general information about published research and is not a prediction. It takes no position on which group is closer to the truth.
FAQ
Did the tournament find that AI will end humanity?
No. Both groups gave a probability below 100 percent. The median forecasts were 0.38 percent and three percent for extinction by 2100.
Who are superforecasters?
In this study they are forecasters who scored highly in previous tournaments on questions that resolve within months or years. The tournament paper uses that record as the basis for comparing them with subject experts.
Why did experts and forecasters not agree after debating?
The authors say they do not know and list it as an open question. The paper notes differences in expected AI timelines and in doubt about large economic effects, but does not claim these explain the split.
Are these numbers still current?
They describe views in 2022. The institute reported that skeptics re-surveyed after new language models appeared had not become noticeably more concerned, but a full repeat of the tournament was not reviewed for this article.
Related reading
- Threat or tool: the wider debate
- Debate topics in AI ethics and impact
- The controversy over AI risks
- Benefits and risks of AI by field
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Existential Risk Persuasion Tournament, Forecasting Research Institute, 2023
- Karger, Rosenberg, Jacobs, Hickman, Hadshar, Gamin, Smith, Williams, McCaslin, Thomas, and Tetlock, Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament, FRI Working Paper #1, July 10, 2023 (revised August 8, 2023)