AI Experts Keep Missing the Mark on Progress Timelines, Forecasting Study Finds
A Forecasting Research Institute analysis shows median expert predictions for AI milestones have consistently lagged actual results, with some benchmarks arriving years early and others still unmet.
Leading artificial intelligence researchers have a poor track record of predicting their own field's progress, according to a new analysis from the Forecasting Research Institute. The study compared expert forecasts against what actually happened, and found the median prediction for several major AI milestones was too conservative by a wide margin. The gap matters because those same experts shape investment decisions, policy debates, and public expectations about what the technology can do.
The clearest miss came at the International Mathematical Olympiad, where AI systems reached gold-medal performance five years earlier than the median expert had projected. Commercial traction also outran predictions: Anthropic's annualized revenue is running at roughly five times the level forecasters expected. Both data points suggest that capability gains and market adoption are compounding faster than the people closest to the work anticipate.
The picture is not uniformly bullish, though. Forecasts for real-world deployments such as self-driving cars have proven more mixed, with promised timelines repeatedly slipping past their targets. That split hints at a pattern worth taking seriously: tasks that can be benchmarked and scored in controlled settings are arriving early, while messy physical-world problems remain stubbornly difficult. The study does not claim experts are always wrong in one direction, only that their errors have clustered on the optimistic side for measurable AI milestones.
For anyone allocating capital or writing rules around AI, the finding cuts against the habit of treating expert timelines as reliable anchors. If median forecasts systematically undershoot, then planning cycles built on those numbers may leave organizations flat-footed when capabilities land ahead of schedule. It also complicates safety and governance work, since regulators often pace their interventions to expected arrival dates that the data suggests are too late.
The Forecasting Research Institute's work adds to a growing body of evidence that AI insiders are not neutral observers of their own field. What remains unanswered is whether the misses reflect genuine uncertainty, institutional incentives to sound cautious, or simply how hard it is to model exponential change. The next test will be whether forecasts for agentic systems and long-horizon reasoning show the same early-arrival pattern, or whether the field finally starts pricing in its own speed.