AI experts consistently underestimated the field's pace, study finds

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

The Forecasting Research Institute published a study, reported on September 24, 2026, finding that prominent AI experts have repeatedly underestimated the pace of AI development. Two specific data points anchor the finding: AI systems reached gold-medal performance at the International Mathematical Olympiad five years ahead of the median expert forecast, and Anthropic's annualized revenue came in at roughly five times the level those same experts had projected. The study compares tracked predictions from named figures in the field against actual outcomes, giving it a more systematic character than anecdotal claims about forecasting failure. Notably, the pattern does not hold uniformly across all domains, as expert predictions about real-world applications such as self-driving cars showed a more mixed record, with some forecasts running ahead of what actually materialized.

The study matters because expert forecasts do not stay in academic journals. They inform hiring decisions, regulatory timelines, investment theses and public policy. If the people closest to the technology have been systematically too conservative about core capability benchmarks, institutions that relied on those forecasts may have prepared for a slower transition than the one now underway. At the same time, the self-driving data complicates any simple narrative: underestimation on research benchmarks does not automatically translate to underestimation on deployment, and the distinction between what AI can do in a controlled setting and what it delivers commercially remains a live question.

None of the three camps, Pro-AI, Anti-AI and Middle Ground, had published reactions at the time of writing. Pro-AI commentators would typically treat this kind of finding as confirmation that caution about AI progress has been misplaced and that continued rapid investment is warranted. Anti-AI voices would likely argue that consistent underestimation by experts is precisely the reason stronger precautionary regulation should not wait for consensus to form around when powerful systems will arrive. The Middle Ground camp would probably point to the mixed forecasting record on deployment as evidence that capability gains and real-world impact are different problems, and that one does not straightforwardly predict the other.

The most useful thing to watch is whether the Forecasting Research Institute releases the underlying dataset of individual predictions, which would allow independent researchers to assess which types of experts, and which types of claims, showed the largest errors. That would help separate a general pattern of underestimation from something more specific to benchmark-style tasks.