Experts and professional forecasters substantially underestimated several recent AI milestones, according to an interim report from the Forecasting Research Institute. Its longitudinal panel includes 339 specialists spanning computer science, industry, economics and policy, alongside forecasters with established prediction records.
The largest gap involved mathematics. AI reached gold-medal performance at the International Mathematical Olympiad in July 2025, five years before the median expert prediction and 10 years before the median superforecaster prediction gathered in 2022. The institute also says systems likely matched a top virology team on a troubleshooting benchmark in April 2025, versus median forecasts of 2030 and 2034. Reported AI-company revenue also rose much faster than panel estimates.
Benchmark gains did not consistently translate into practical impact. In one controlled biology trial, 5.2 percent of participants using a language model completed laboratory tasks, compared with 6.6 percent using the internet alone. The small study found no measurable benefit. Forecasts for autonomous ride-hailing may also have been too optimistic.
The institute cautions that its interim method is asymmetrical: an underestimate becomes visible as soon as an event occurs, while an overestimate can be identified only after its deadline passes. Some assessments also rely on model projections using newer information. The findings therefore support faster forecast updates, not a blanket conclusion that every AI outcome arrives sooner than expected.