Current estimates of AI-caused existential catastrophe are too assumption-dependent and weakly grounded in evidence to serve as standalone policy evidence. That does not make probabilities useless: policymakers can use them as conditional inputs when the outcome, time horizon, scenario and evidential basis are explicit—and when decisions are tested across a range of plausible estimates.
What does an AI existential-risk probability actually measure?
A probability is only interpretable if the event and time frame are clear. “How likely is AI to cause human extinction?” is not the same question as whether AI could cause an unrecoverable societal collapse, or whether it could contribute to a large death toll. Forecasts using different endpoints cannot be compared as though they estimate the same risk.
As an Amazon Associate I earn from qualifying purchases.
The Forecasting Research Institute (FRI) uses the question “Will AI cause an existential catastrophe by 2100?” Its definition includes extinction or specified forms of unrecoverable collapse. LEAP Wave 9 instead asks about a “global AI-related catastrophe,” defined as more than 10% of the population alive at the beginning of a five-year period dying by its end. That is a severe outcome, but it is not a forecast of human extinction.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →The horizon matters too. A forecast for 2100 says nothing directly about the chance of catastrophe by 2030, and estimates conditional on rapid AI progress answer a different question from unconditional estimates. A number without those qualifications can sound more definite than the underlying forecast is.
#1 Best Overall
What do the recent forecasts say?
Two recent FRI projects illustrate both the range of estimates and why their questions must stay attached to their numbers.
FRI’s 2024 adversarial collaboration
FRI brought together 22 selected participants: 11 “AI skeptic” participants—nine superforecasters and two domain experts—and 11 AI-concerned domain experts. After several weeks of reviewing evidence and forecasting together, the groups’ median estimates for AI-caused existential catastrophe by 2100 remained far apart.
Rank #2
| Group | Median at beginning | Median at end |
|---|---|---|
| AI skeptic group | 0.10% | 0.12% |
| AI-concerned group | 25% | 20% |
The small change in these medians does not establish that either group is right. The participants were selected for the collaboration, not sampled to represent all experts or the public. The project is evidence of persistent disagreement among those participants, not a measurement of the true probability.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteLEAP Wave 9 in 2026
LEAP Wave 9 collected responses from May 19 to June 10, 2026, and released its report on June 30. It included 194 experts, 53 superforecasters and 612 public respondents. For its defined global AI-related catastrophe outcome, the median expert forecast by 2100 was 2% under slow progress and 10% under rapid progress.
Those are scenario-conditioned forecasts of an outcome involving more than 10% of a population dying within five years; they are not estimates of extinction, and they are not historical event rates. The gap between the slow- and rapid-progress scenarios is a reminder that assumptions about AI development can materially affect the result.
Why do informed people disagree so much?
FRI’s collaboration found disagreement not just about a single numerical input, but about several links in the chain from AI development to catastrophe. Participants differed over capability timelines, whether AI would develop goals connected to extinction, how difficult human extinction would be, and how societies would respond. The report also points to broader differences in worldview. Short-term indicators examined in the project explained only a modest share of the forecast gap.
Working together and summarizing one another’s arguments did not produce substantial convergence. That is useful evidence about the difficulty of the question, but it is not a test showing that one side’s estimates are better calibrated or more accurate.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsWhy a probability can be a poor summary of uncertainty
Many forecasts are judgments about events that have not occurred, rather than estimates derived from a large record of similar past events. For long-horizon AI catastrophe, the available evidence and detailed theory are sparse. There is no established track record here showing that century-scale existential-risk forecasts are well calibrated.
Best Value
CSET’s May 2026 brief distinguishes uncertainty from randomness. Randomness is the variation in outcomes one might model with probabilities; ignorance reflects missing evidence or limited understanding of how a scenario would unfold. Andrew Lohn, the brief’s author, writes: “In AI risk, rather than in dice rolls, ignorance is the dominant form of uncertainty, not randomness, so the best techniques are not always probabilistic.”
The brief does not argue that probabilities are always inappropriate. It describes belief and plausibility as alternative ways to ask how strongly the available evidence supports or argues against a scenario. These approaches can make the evidential limits more visible instead of compressing uncertainty into a single point estimate.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What do these studies establish—and what do they not?
- They document judgments, not observed rates. FRI’s collaboration and LEAP’s panel record what selected participants forecast under specified questions and conditions. The figures are not counts of past AI-caused catastrophes.
- They show disagreement and sensitivity to assumptions. The FRI groups stayed far apart, while LEAP’s expert median differed between its slow- and rapid-progress scenarios.
- They do not establish which estimate is correct. The reviewed studies do not settle the true probability of an AI-caused existential catastrophe, demonstrate long-horizon calibration, or measure whether policies selected using these probabilities improve outcomes.
A separate 2025 preprint by Severin Field surveyed 111 AI experts. In that study, 78% agreed or strongly agreed that technical AI researchers should be concerned about catastrophic risks, and 21% said they had heard of instrumental convergence. Those results describe the respondents’ views and familiarity with a concept; they do not test forecast accuracy or validate a particular risk estimate.
Recommended Free Tools
How should policymakers use the numbers?
A probability can help structure a decision, but it should not carry the decision by itself. The useful question is not simply “What is the number?” It is “What action would this estimate change, and does that action remain sensible if the estimate is substantially higher or lower?”
- Name the outcome. Say whether the policy aims to reduce extinction risk, prevent unrecoverable collapse, limit catastrophic misuse, preserve human control, or address another defined harm.
- State the horizon and conditions. Distinguish a near-term forecast from one to 2100, and unconditional judgments from estimates conditional on slow or rapid progress.
- Show the basis and uncertainty. Identify whether the number comes from expert judgment, a panel, a model, observed data or theoretical argument. Make important unknowns and evidence on both sides visible.
- Compare decisions across plausible probabilities. If a precaution remains worthwhile over a broad range of estimates, policymakers may not need to select one contested point estimate first. If the preferred action changes sharply with the number, the assumptions behind that sensitivity deserve attention.
- Track observable indicators. Specify what developments would strengthen, weaken or otherwise change the case for an intervention, so a forecast can be revisited as evidence changes.
These are decision-analysis practices, not outcomes proven by the cited forecasting projects. Their purpose is to keep a contested estimate in its proper role: one input among the event definition, evidence, uncertainty, costs of action and consequences of waiting.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
