# Top AI Experts Badly Underestimated How Fast the Field Is Moving, Study Finds
The Forecasting Research Institute has documented a troubling pattern: leading AI researchers have systematically underestimated the speed of progress in their own field. This matters because expert forecasts shape policy, investment, and public expectations around AI development.
The gap between predictions and reality is substantial. AI reached gold-medal level at the International Mathematical Olympiad five years faster than the median expert forecast suggested. This wasn't a close miss. Experts looked at mathematical reasoning benchmarks and said the timeline would be longer. It wasn't.
Anthropic's revenue tells a similar story. The AI safety-focused company reached revenue levels about five times higher than what forecasters predicted. This signals both rapid commercial adoption and experts' blindness to how quickly AI capabilities could translate into business value.
The pattern extends beyond specific metrics. Experts have underestimated both the rate of capability improvements and the speed at which those improvements found commercial applications. When researchers predicted when AI would solve certain tasks, they built in too much slack. Reality moved faster.
This creates a real problem for institutions that rely on expert judgment. Governments designing AI regulation base timelines on expert testimony. Investors decide where to allocate capital based on forecasted adoption curves. Companies plan product roadmaps around expert consensus on capability emergence. When experts are systematically wrong in one direction, downstream decisions compound the error.
The study does find a more complicated picture for real-world deployment. Self-driving cars present a mixed forecast record. Some experts overestimated timeline progress here, while others underestimated. This suggests the challenge isn't uniform. Benchmarks and narrow tasks show clear underestimation. Real-world applications with complex edge cases show scattered accuracy.
Why the systematic underestimation? Several factors likely contribute. Experts often anchor to previous technology curves. AI progress has followed a different trajectory than prior computing revolutions. The combination of scale, data, and architectural innovation created non-linear acceleration that forecasters didn't fully price in. Experts also tend toward conservatism in public statements, especially on contentious topics. Nobody wants to sound like they're hyping technology.
There's also the problem of expertise itself. The researchers closest to cutting-edge AI development work often see slow, incremental progress in their daily work. This granular view can obscure the broader acceleration pattern. You see your own difficulty with a problem. You don't see how a different architecture or scale might make that problem trivial.
The study's timing matters. As AI regulatory frameworks take shape globally, this data suggests policymakers should treat expert timelines skeptically. If experts have underestimated progress by years on multiple fronts, regulations built on expert testimony about when certain capabilities emerge could prove rapidly obsolete.
This also reshapes how the field should approach forecasting going forward. The Forecasting Research Institute's work suggests AI development needs dedicated forecasting infrastructure separate from the people building the technology. Just as financial markets employ dedicated forecasters, AI governance needs people whose job is specifically to track and predict capability emergence, separate from researchers pushing the boundaries.
The lesson cuts both ways. Underestimation creates risk. But it also suggests skepticism toward doomsday timelines based on expert consensus. If experts missed upside by years, they might also be wrong about when certain risks emerge.