Study finds AI experts underestimated progress on benchmarks and adoption

Not all forecasts were overly conservative. Biosecurity and virology experts’ predictions about AI-assisted lab tasks were closer to reality, though a small trial showed minimal impact from language models. Self-driving car forecasts were also revised downward, with experts’ 7.3% estimate for autonomous ride-hailing trips in 2027 contradicted by an LLM projection of 2.5%. FRI notes that forecasts on economic growth, employment, and AI harms remain unreliable. Meanwhile, respondents’ views on AI’s long-term significance have shifted upward, with the probability of it becoming a ‘technology of the century’ rising from 31% to 36% for experts and 28% to 35% for superforecasters over nine months. FRI plans to refine its methods, including tracking rapid-progress forecasters and comparing LLM projections with human predictions.