How much risk of catastrophe?
Experts and superforecasters give an AI-caused catastrophe before 2100 probabilities that differ by a factor of five. What was asked, what they answered and why nobody can measure it.
What counts as a catastrophe here
We call a catastrophe an event that kills at least 10% of humanity within a 5-year period. It is the definition used by the forecasting tournament cited below and by our futures model. Extinction is a different, far more extreme question.
What forecasters say
Between June and October 2022, before ChatGPT was released, a tournament brought together experts in each risk and superforecasters, people with a proven track record of accuracy, to estimate long-term risks. For a catastrophe from any cause before 2100 (pandemics, nuclear war, AI and the rest), the superforecasters’ median was 9.05% and the experts’ 20%. For a catastrophe caused by AI alone, the AI experts’ median was 12% and the superforecasters’ 2.13%. For AI-caused extinction, 3% and 0.38%.Source 1
On AI, the experts gave five to eight times more probability than the superforecasters.
What researchers say
A 2023 survey of 2,778 researchers who publish at the top AI conferences found that between 38% and 51% gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction.Source 2
What the international report says
The 2026 International AI Safety Report, commissioned by more than 30 countries, concludes that current systems lack the capabilities to pose loss-of-control risks, but that they are improving in related areas such as operating autonomously. It warns that evidence on these risks is slow to emerge and hard to assess, and that some models are getting better at telling when they are being tested.Source 3
In the model
Our model computes the cumulative risk of catastrophe from any cause, not just AI: pandemics, wars and natural disasters too. Keeping Pause below 9% in 2100 and Race above 20% is an editorial choice, anchored in the medians for a catastrophe from any cause: the superforecasters’ (9.05%) and the experts’ (20%). At the centre of the matrix, the parameter that moves the figure most is the risk added by power concentration, calibrated against these estimates rather than measured, as the sensitivity analysis shows. In the futures with a lot of peace, the baseline risk weighs more.
What we do not know
These figures are judgments, not measurements. Nobody has seen a global catastrophe caused by AI, so there is no frequency to count. That serious groups disagree so much is perhaps the most honest data point: the uncertainty is enormous.
Sources
- Karger et al., Forecasting Research Institute (2023). Forecasting Existential Risk: Evidence from a Long-Run Forecasting Tournament.
- Grace et al. (2024). Thousands of AI Authors on the Future of AI.
- Bengio et al., commissioned by more than 30 countries (2026). International AI Safety Report 2026.