FRI Study Finds Experts and Superforecasters Systematically Underestimated AI Capabilities and Adoption

The Forecasting Research Institute (FRI) reviewed AI progress forecasts from mid-2022 to August 2026, finding that both domain experts and superforecasters significantly underestimated the speed of AI advancement. Key findings include:

• Benchmark Surprises: AI models achieved International Mathematical Olympiad (IMO) gold-level performance in July 2025, five years earlier than the median expert prediction (2030) and ten years earlier than the median superforecaster prediction (2035). Similarly, AI models likely matched top virologist teams on the VCT benchmark in April 2025, years ahead of the 2030–2034 forecasts.
• Revenue Underestimation: Forecasts for AI company revenue were drastically low. For instance, the combined annualized revenue run rate of Anthropic and OpenAI was forecasted at $70–90 billion for the end of 2026, while current reports indicate a value closer to $140 billion.
• Capability vs. Utility: While benchmark capabilities were underestimated, forecasters overestimated the immediate practical utility of AI in specific non-benchmark tasks. In a biorisk study, experts predicted 22.5% of STEM undergraduates could complete complex lab tasks with LLM assistance, but only 5.2% succeeded.
• Methodological Bias: FRI notes that their methodology is biased toward identifying underestimates, as it is easier to spot when current values exceed median forecasts than to confirm overestimates before resolution dates pass.