Study finds AI experts underestimated the pace of AI
The Forecasting Research Institute says its panel put a maths-olympiad gold two years late, and top AI revenue at a fifth of the reported figure.
Experts on a forecasting panel put AI milestones later than they arrived, according to the Forecasting Research Institute, in work reported on Thursday by The Decoder. The panel put gold-medal performance at the International Mathematical Olympiad in 2027. It happened in July 2025.
The institute drew on 339 experts in the first round of its Longitudinal Expert AI Panel: 76 computer scientists, 76 industry experts, 68 economists and 119 policy specialists, alongside superforecasters picked for past accuracy. Thirty of the computer scientists are professors at top-20 institutions, according to the report.
Revenue was the widest gap. The panel's forecast for the highest annual recurring revenue at the end of 2026 was $20 billion. Anthropic is reported to have reached about $100 billion by this month. Superforecasters were later than the experts on the olympiad, putting it at 2032.
The pattern does not hold everywhere. On self-driving, the median expert put autonomous ride-hailing at 7.3% of United States trips in 2027, above the 2.5% the institute derived from a language model. On virology, experts said 2030 and superforecasters 2034; the institute dates the moment to April 2025.