AI Progress Outpaces Expert Forecasts: FRI Study Reveals Significant Underestimations

The Forecasting Research Institute (FRI) published an interim analysis of its LEAP forecasts, revealing that senior experts and superforecasters systematically underestimated the rapid pace of AI progress, with AI reaching gold-medal level at the International Mathematical Olympiad in July 2025, five years before the median expert forecast. For broader context, explore our AI News.

Expert Forecasts Miss Key Milestones

The FRI's LEAP initiative gathered predictions from 339 experts, including computer scientists, industry specialists, economists, and AI policy experts. The interim analysis indicates that many significant AI advancements occurred years ahead of these expert projections. For instance, AI achieved gold-medal level performance at the International Mathematical Olympiad in July 2025, a full five years before the median expert forecast and ten years earlier than the median superforecaster prediction. Experts had assigned an average probability of only 24.6 percent to this outcome, while superforecasters gave it a mere 9.7 percent.

Underestimated Capabilities Across Domains

Similar patterns of underestimation were observed in other critical areas. The Virology Capabilities Test milestone, for example, was likely reached in April 2025. This significantly predates the median expert forecast of 2030 and the median superforecaster forecast of 2034. These findings suggest a broader trend where the rate of AI development consistently exceeds even the most informed human expectations.

AI Company Revenue Surpasses Projections

The financial growth of AI companies also significantly outstripped expert predictions. Median forecasts for the highest AI-company annual recurring revenue by the end of 2026 were $20 billion from experts, $16 billion from economists, and $25 billion from superforecasters. However, Anthropic, a prominent AI company, reported approximately $100 billion in revenue by September 2026, far exceeding all these projections. This substantial difference underscores the difficulty in accurately predicting the economic impact and market adoption of rapidly evolving AI technologies.

Not All Predictions Were Underestimated

While many areas saw AI progress accelerate beyond forecasts, not all predictions followed this trend. For example, only 5.2 percent of participants successfully completed biological lab tasks using a large language model (LLM) with internet access. This contrasts sharply with the 40 percent expected by virologists, indicating that certain complex, real-world applications of AI may still present significant challenges or require more specialized integration than anticipated.

The Future of AI Forecasting

The Forecasting Research Institute plans to continuously update its LLM forecasts alongside human predictions. This approach acknowledges that some AI models are already demonstrating forecasting capabilities that match or even surpass superforecasters on specific types of questions. Integrating AI-driven forecasts could provide more dynamic and accurate insights into future AI progress, potentially mitigating the systematic underestimation observed in human expert predictions.

Conclusion

The interim analysis from the Forecasting Research Institute's LEAP study provides clear evidence that the pace of AI development has consistently outstripped expert and superforecaster predictions across multiple critical benchmarks. From achieving gold-medal status in mathematical olympiads to exceeding revenue forecasts, AI's rapid advancement presents both challenges and opportunities for understanding and guiding its future trajectory. The FRI's commitment to integrating AI-powered forecasts suggests a future where our understanding of AI's potential may become more aligned with its actual progress.

Sources

• How Accurate Have AI Progress Forecasts Been So Far?
• Top AI experts badly underestimated how fast the field is moving, study finds