How fast is AI improving? The question usually goes to university experts, heavily cited researchers and economists. An interim report from the Forecasting Research Institute (FRI) compared their past forecasts with what actually happened, and the finding is that the forecasts mostly ran low.

Who is on the panel

FRI has been collecting forecasts on AI progress since mid-2022. The first round of its longitudinal panel drew 339 experts:

  • 76 computer scientists (30 of them professors at top-20 institutions, 10 among the 200 most-cited authors),
  • 76 industry experts,
  • 68 economists,
  • 119 AI policy specialists.

The panels also include "superforecasters": generalists chosen for a proven record of accurate predictions rather than domain expertise.

The widest gap: mathematics

AI reached gold-medal level at the International Mathematical Olympiad in July 2025. That is five years ahead of the median expert forecast and ten years ahead of the median superforecaster forecast. Those predictions were gathered in 2022, before ChatGPT; FRI says the pattern held afterwards too.

Other examples:

  • Millennium problem: in a survey in late summer 2025, experts put the median odds of such a problem being solved by the end of 2027 at 10%, superforecasters at 5.4%.
  • Virology: experts said models would not match a top team of virologists on a troubleshooting benchmark until 2030; superforecasters said 2034. FRI says it likely happened in April 2025.
  • Revenue: for the highest annual recurring revenue of any AI company at the end of 2026, experts said $20 billion, economists $16 billion and superforecasters $25 billion. FRI points to a figure of roughly $100 billion as of September 2026.

Not every forecast ran low

This is the part not to skip. In some areas the forecasts were too high:

  • In biosecurity, only 5.2% of participants completed biological tasks using language models; experts had expected 22.5% and virologists 40%.
  • Autonomous vehicles show similar over-optimism: experts forecast 7.3% of US ride-hailing trips would be autonomous by 2027, against a modelled 2.5%.

So the error is not one-directional. Capabilities measured on benchmarks arrived faster than expected; the spread of those capabilities into the real world has been slower.

The report's own caveat

FRI states a methodological limit itself: underestimates become obvious the moment reality overtakes them, while overestimates only become clear once a deadline passes. That tilts this kind of review towards finding that experts were too conservative.

The practical takeaway: treat forecasts about when a capability will arrive with caution, but keep the distinction between "this capability exists" and "this capability changed my work". By this report, the first arrives faster than expected and the second more slowly.