AI Researchers’ 2024 Forecasts: High Risks and Diverging Timelines

Summary

AI Impacts summarizes the fourth edition of its recurring survey of AI researchers, conducted in December 2024. The survey received 1,580 valid responses from researchers associated with six top-tier AI venues, although each respondent answered a randomized subset of questions and the authors acknowledge likely non-response bias. On average, respondents assigned an 18% probability to future AI advances causing human extinction or similarly permanent and severe disempowerment; more than half assigned at least 10%, and one-third assigned at least 20%. Researchers educated in Asia gave higher estimates than those in the United States or Europe, while a China-specific result remains preliminary because of the small sample and possible effects of VPN use and response rates. Respondents also reported concern about multiple harmful scenarios and expected most traits associated with dangerous AI capabilities to be more common 20 years from now than in the previous survey. They were pessimistic that users would know the true reasons behind typical state-of-the-art systems’ decisions by 2029. Estimates of the importance of research inputs were highly dispersed: on average, halving hardware, data, or other inputs over the preceding decade was seen as reducing progress by roughly half, but individual views varied widely. Forecasts for narrow abilities were comparatively short, with all but five tasks expected within ten years on average. Estimates for broadly human-level AI diverged sharply depending on the question: 16 years for a task-based measure versus 72 years for an occupation-based measure. Both timelines fell substantially from the prior survey, by five and 17 years respectively, and the task-based forecast put a 10% chance on achievement by 2027. Researchers were roughly divided over whether a faster, slower, or unchanged pace of global AI research would be most desirable.