FRI's review of AI forecast accuracy

The Forecasting Research Institute reviewed four years of expert AI forecasts and found that experts and superforecasters consistently underestimated AI capabilities gains and lab revenues, while overestimating biorisk uplift from AI. Predictions about diffusion rates were more mixed. FRI drew on multiple studies including the Existential Risk Persuasion Tournament, an AI adversarial collaboration, an LLM-biorisk study, an AI-cyber risk pilot, and the Longitudinal Expert AI Panel. On specific forecasts, superforecasters gave only a 1.7% probability to a Millennium Prize being solved as soon as it was, while experts gave 4.5% and placed a median prediction at 2040. Self-driving cars remain a counterexample, representing roughly 0.6% of ride-share trips, slower diffusion than many predicted. FRI also carries a significant methodological caveat: its approach is structurally biased toward surfacing underestimates, because a median forecast is revealed as too low the moment real-world performance crosses it, whereas an overestimate can only be confirmed after the full resolution date passes. FRI further notes that the vast majority of its AI progress forecasts remain unresolved, so the overall accuracy picture could shift substantially. Separately, AI models have surpassed top human forecasters on MetaculusBench, and an AI reportedly won the Summer 2026 Metaculus Cup, beating every human participant.