Specialized AI Forecasters Match 68.8 Superforecaster Score, Nearing Humans at 68.9
Updated
Updated · Financial Times · Sep 16
Specialized AI Forecasters Match 68.8 Superforecaster Score, Nearing Humans at 68.9
3 articles · Updated · Financial Times · Sep 16
Summary
ForecastBench put the best specialized AI forecasters at 68.8 on its accuracy index, essentially level with human superforecasters at 68.9 and ahead of ordinary human forecasts at 62.8.
The benchmark scores probabilistic predictions on a 0-100 scale, where 50 is equivalent to constant 50-50 guesses and 100 represents perfect confidence paired with perfect accuracy.
ForecastBench cautioned the comparison is not like-for-like: human superforecaster results came from 2024 questions, while AI systems are answering a different set today, with statistical adjustments becoming less reliable over time.
Regular untuned LLMs already score above 61, suggesting broad gains in machine forecasting even before specialization, though the report says both humans and generic models remain far closer to guesswork than perfection.
The article argues the bigger issue is not just accuracy but power and process: AI forecasts could shape outcomes, while replacing human forecasting may forfeit the deliberation and depolarizing benefits of doing it ourselves.