AI Progress Is Outpacing Top Expert Predictions by Years, FRI Study Finds

Forecasting Panel Tracks Massive Gaps Between Predictions and Reality

Leading artificial intelligence specialists, economists, and professional forecasters have significantly underestimated the pace of technical and financial breakthroughs across the AI sector, according to an interim report released by the Forecasting Research Institute (FRI). Drawing on surveys compiled since mid-2022, the findings compare forward-looking predictions against real-world milestones reached over the past several years.

The evaluation centers on FRI's Longitudinal Expert AI Panel (LEAP), which surveyed 339 specialists, including 76 computer scientists, 76 industry professionals, 68 economists, and 119 AI policy analysts. The computer science cohort included 30 professors from top-20 universities and 10 of the world's 200 most-cited AI researchers. Their projections were assessed alongside those of superforecasters—generalist predictors with documented track records of forecasting accuracy. Across both groups, median projections consistently placed major capability leaps years later than they actually arrived.

Mathematics, Science, and Revenue Dramatically Exceed Forecasts

The starkest divergence emerged in advanced mathematics. In July 2025, artificial intelligence systems achieved gold-medal performance at the International Mathematical Olympiad. That milestone arrived five years ahead of the median expert forecast and a full decade earlier than superforecasters expected. When surveyed in 2022, experts gave gold-medal performance only an 8.6 percent chance of occurring when it did, while superforecasters assigned it a 2.3 percent likelihood. Across all benchmark achievements that have materialized, experts gave an average probability of 24.6 percent, compared to 9.7 percent from superforecasters.

Other high-level technical benchmarks saw similar discrepancies:

• Millennium Prize Problem: Following indications that AI may have solved a Millennium Prize Problem (with official verification criteria still under evaluation), FRI noted that August and September 2025 surveys had placed the median odds of an AI solving such a problem by late 2027 at just 10 percent among experts and 5.4 percent among superforecasters.
• Virology and Cybersecurity: When asked when AI systems would match elite teams on the Virology Capabilities Test troubleshooting benchmark, the median expert timeline pointed to 2030, and superforecasters projected 2034. FRI found this threshold was likely met around April 2025. Parallel underestimates occurred on cybersecurity benchmarks.
• Financial Growth: Forecasters were unprepared for commercial expansion. Experts predicted the highest annual recurring revenue (ARR) for any single AI firm by late 2026 would reach a median of $20 billion, with economists at $16 billion and superforecasters at $25 billion. Reported figures outpaced those expectations, reaching $65 billion by July 2026, while Anthropic reached an estimated ARR of roughly $100 billion by September 2026.

Predictions did not fall short across every category. Forecasters overshot the immediate impact of AI on wet-lab biology tasks: virologists had predicted a 40 percent task-completion rate with AI assistance, but a small controlled trial showed human participants completed only 5.2 percent of tasks with an LLM and web access—slightly lower than the 6.6 percent success rate of subjects using internet search alone. Forecasters also anticipated faster autonomous vehicle penetration, predicting driverless vehicles would capture 7.3 percent of U.S. ride-hailing trips by 2027, compared to LLM-assisted projections that place the number at 2.5 percent.

Rising Expectations and Faster Tracking Methods

As technical progress has compressed timelines, surveyed specialists have adjusted their outlooks upward. Over a nine-month period, the average probability assigned by experts to AI becoming a "technology of the century"—on par with the historical impact of electricity—rose from 31 percent to 36 percent. Superforecasters lifted their estimate from 28 percent to 35 percent over the same window.

To keep pace with the compressed development cycle, FRI plans to highlight a dedicated cohort of panelists expecting aggressive advancement through 2040 and deploy continuously updated model-generated forecasts. The institute noted that automated systems evaluated on ForecastBench already rival superforecasters on specific question categories.

FRI also highlighted an inherent structural limitation in its interim findings: while underestimates become obvious the moment a model achieves a milestone ahead of schedule, overestimates remain unconfirmed until deadlines expire, naturally weighting early evaluations toward examples of underprediction.