Thousands of AI Authors on the Future of AI

TL;DR

This paper reports ESPAI 2023, the largest AI expert forecast survey to date, assessing how quickly AI will progress on a broad set of tasks and the potential social impacts. Using framing variations and gamma distribution aggregation across 2778 researchers from six venues, it finds many milestones could be feasible within a decade, and HLMI and FAOL forecasts shift earlier since 2022, albeit with substantial uncertainty. The results reveal meaningful concern about extreme AI risks and a strong push to prioritize AI safety and alignment research, while highlighting framing and sample limitations that affect interpretability. Together, the findings provide a dense, data driven snapshot of expert views on AI futures, with implications for policy, governance, and risk management.

Abstract

In the largest survey of its kind, 2,778 researchers who had published in top-tier artificial intelligence (AI) venues gave predictions on the pace of AI progress and the nature and impacts of advanced AI systems The aggregate forecasts give at least a 50% chance of AI systems achieving several milestones by 2028, including autonomously constructing a payment processing site from scratch, creating a song indistinguishable from a new song by a popular musician, and autonomously downloading and fine-tuning a large language model. If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047. The latter estimate is 13 years earlier than that reached in a similar survey we conducted only one year earlier [Grace et al., 2022]. However, the chance of all human occupations becoming fully automatable was forecast to reach 10% by 2037, and 50% as late as 2116 (compared to 2164 in the 2022 survey). Most respondents expressed substantial uncertainty about the long-term value of AI progress: While 68.3% thought good outcomes from superhuman AI are more likely than bad, of these net optimists 48% gave at least a 5% chance of extremely bad outcomes such as human extinction, and 59% of net pessimists gave 5% or more to extremely good outcomes. Between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction. More than half suggested that "substantial" or "extreme" concern is warranted about six different AI-related scenarios, including misinformation, authoritarian control, and inequality. There was disagreement about whether faster or slower AI progress would be better for the future of humanity. However, there was broad agreement that research aimed at minimizing potential risks from AI systems ought to be prioritized more.

Figures (20)

• Figure 1: Most milestones are predicted to have better than even odds of happening within the next ten years, though with a wide range of plausible dates. The figure shows aggregate distributions over when selected milestones are expected, including 39 tasks, four occupations, and two measures of general human-level performance (see Section \ref{['sec:hlmi']}), shown as solid circles, open circles, and solid squares respectively. Circles/squares represent the year where the aggregate distribution gives a milestone a 50% chance of being met, and intervals represent the range of years between 25% and 75% probability. Note that these intervals represent an aggregate of uncertainty expressed by participants, not estimation uncertainty. The displayed milestone descriptions are summaries; for full descriptions, see Appendix \ref{['app:milestone_descriptions']}.
• Figure 2: Expected feasibility of many AI milestones moved substantially earlier in the course of one year (between 2022 and 2023). The milestones are sorted (within each scale-adjusted chart) by size of drop from 2022 forecast to 2023 forecast, with the largest change first. The year when the aggregate distribution gives a milestone a 50% chance of being met is represented by solid circles, open circles, and solid squares for tasks, occupations, and general human-level performance respectively. The three groups of questions have different formats that may also influence answers. For full descriptions of the summarized milestones, see Appendix \ref{['app:milestone_descriptions']}.
• Figure 3: Aggregate forecast for 50th percentile arrival time of High-Level Machine intelligence (HLMI) dropped by 13 years between 2022 and 2023. The forecast for 50th percentile arrival time of Full Automation of Labor (FAOL) dropped by 48 years in the same period. However, there is still a lot of uncertainty. "Aggregate Forecast" is the mean distribution over all individual cumulative distribution functions. For comparison, we included the 2022 Aggregate Forecast. To give a sense of the range of responses, we included random subsets of individual 2023 and 2022 forecasts. Note that the thinner 'confidence interval' in 2023 (compared to 2022) is due to our increased confidence about the average respondents' views due to a larger sample size, not respondents' predictions converging.
• Figure 4: Most respondents indicated that the pace of progress in their area of AI increased between the first and second half of their time in a field. Participants were asked whether the second half of the time they had spent working in their area of AI saw more progress than the first half. The median time working in the area was 5 years.
• Figure 5: Estimated reduction in AI progress if inputs had been halved over the past decade. Red dots represent means. Boxes contain the 25th to 75th percentile range; middle lines are medians. Whiskers are the least and greatest values that are not more than 1.5 times the interquartile range from the median. Participants estimated that halving the drop in costs of computing would have had the greatest effect on AI progress over the last decade, while halving 'researcher effort' and 'progress in AI algorithms' would have had the least effect. Overall, all the included inputs were seen as having contributed substantially to AI progress.