AI Progress Forecasts: Experts Underestimated the Pace

Predicting how fast artificial intelligence will improve usually falls to a familiar group: professors at leading universities, heavily cited AI researchers and seasoned economists. They are the people policymakers and businesses turn to when they want a sense of what comes next.

According to a detailed interim report from the Forecasting Research Institute (FRI), that group has often been too cautious. On benchmarks and on some adoption indicators, progress over the past few years has run well ahead of what experts expected.

Who was asked: a panel built from the top of the field

The FRI has been gathering forecasts on AI progress since mid-2022 through several studies and projects. The first round of its Longitudinal Expert AI Panel, or LEAP, included 339 experts: 76 computer scientists, 76 industry specialists, 68 economists and 119 AI policy professionals. Among the computer scientists were 30 professors at top-20 institutions and 10 of the 200 most-cited AI authors.

The institute also brought in superforecasters. These are generalists without deep AI expertise, but with a proven record of making accurate predictions. Having both groups gives a useful comparison: domain knowledge on one side, forecasting discipline on the other.

Milestones that arrived years ahead of schedule

The clearest miss involves the International Mathematical Olympiad. AI reached gold-medal level in July 2025. That was five years earlier than the expert median forecast and ten years earlier than the superforecaster median. Those predictions were collected in 2022, before ChatGPT was released, so some surprise is understandable. The pattern, however, continued after that point.

A possible second case concerns the Millennium Prize Problems. In a survey held in August and September 2025, experts gave a median probability of 10 percent that AI would solve one by the end of 2027. Superforecasters put it at 5.4 percent. OpenAI may already have solved the Navier-Stokes problem with an internal model, although it is not yet clear whether that result meets the criteria used to resolve the question.

Virology shows a similar gap. In one study, experts expected AI models to match a top team of virologists on a troubleshooting benchmark by 2030, while superforecasters expected 2034. The FRI says this probably happened as early as April 2025. A cybersecurity benchmark produced comparable underestimates.

The business numbers point the same way. Asked about the highest annual recurring revenue of any AI company by the end of 2026, experts gave a median of $20 billion, economists $16 billion and superforecasters $25 billion. As a comparison value that has likely already been reached, the FRI cites roughly $100 billion at Anthropic for September 2026.

Real-world use: where experts were too optimistic

The record is not one-sided. In one study, biosecurity experts estimated that 22.5 percent of participants would complete biological lab tasks with help from a language model. Virologists expected 40 percent and superforecasters 16.2 percent. In the controlled experiment, only 5.2 percent succeeded with a language model and internet access, compared with 6.6 percent who had internet access alone. No advantage from the model was visible, though the experiment was small. It is a useful counterpoint to reports of AI making progress in biology labs, where the setup is very different from a novice being guided by a chatbot.

Self-driving cars may be another area where expectations ran high. Experts gave a median of 7.3 percent for the share of US ride-hailing trips that would be autonomous in 2027. An LLM-based projection puts the figure at 2.5 percent. On economic growth, employment and major AI-related harms, the FRI says it is still too early to judge whether the forecasts hold up.

Meanwhile, the forecasters themselves are adjusting. Among people who took part in both surveys, the average probability that AI would be the "technology of the century" rose within nine months from 31 to 36 percent for experts, and from 28 to 35 percent for superforecasters.

What changes next: faster methods and a note of caution

The FRI plans several changes to its approach. It will highlight a subgroup of respondents who expect very rapid AI progress by 2040. It will also publish continuously updated LLM forecasts alongside the human ones. According to ForecastBench, some models now match superforecasters on certain types of questions. Once enough data is available, the institute also wants to identify the most accurate LEAP participants and give their forecasts more weight.

The institute is open about one important limitation. An underestimate becomes visible as soon as an observed value overtakes a forecast. An overestimate can only be confirmed once the deadline has passed. That means an interim review is structurally tilted towards finding underestimates. In addition, some of the preliminary assessments rely on LLM projections, which are speculative in their own right and draw on information that became available later.

For anyone planning around AI, the takeaway is practical rather than dramatic. Even the best-qualified forecasters have struggled to keep pace with capability gains on paper, while real-world adoption has sometimes moved more slowly than expected. Both errors matter, and the gap between them is where most of the uncertainty now sits.