FRI Study Finds Experts Underestimated AI Progress
A new report by the Forecasting Research Institute reveals that top computer scientists and economists have consistently underestimated how quickly artificial intelligence is advancing.
The Forecasting Research Institute (FRI) released an interim report analyzing predictions from its Longitudinal Expert AI Panel (LEAP), which includes 339 specialists: 76 computer scientists, 76 industry experts, 68 economists, and 119 policy experts. The panel features 30 professors from top-20 universities and 10 of the 200 most-cited AI authors, alongside superforecasters. The study found these specialists repeatedly underestimated AI milestones. For instance, AI achieved gold-medal performance at the International Mathematical Olympiad in July 2025, beating the median expert prediction by five years and the superforecaster prediction by ten years. Experts gave this milestone a 24.6 percent probability, while superforecasters estimated it at 9.7 percent.
Other benchmarks showed similar gaps. In virology, experts predicted AI would not match a top team on the Virology Capabilities Test until 2030, and superforecasters estimated 2034, but the milestone likely occurred in April 2025. AI also potentially solved a Millennium Prize Problem, an event experts gave only a 10 percent chance of happening by late 2027. On the economic front, experts predicted the highest annual recurring revenue for any AI firm would reach $20 billion by the end of 2026. However, Anthropic reportedly neared $100 billion in September 2026, and OpenAI surpassed $65 billion in July 2026.
Not all forecasts were too conservative. In a biosecurity trial, only 5.2 percent of participants using a language model completed biological lab tasks, compared to 6.6 percent using only the internet, falling short of the 22.5 percent predicted by experts. Autonomous driving forecasts also appeared high, with experts predicting a 7.3 percent share of US ride-hailing trips by 2027, compared to an LLM projection of 2.5 percent.
For industry practitioners, these findings suggest that relying on traditional expert consensus to plan product roadmaps may leave organizations flat-footed. Because capabilities are outstripping predictions, developers must prepare for rapid, disruptive breakthroughs rather than gradual linear progress. To adapt, FRI is now integrating continuous LLM-based forecasts, utilizing platforms like ForecastBench, to keep pace with the accelerating field.
This is our own summary of reporting by The Decoder
