Forecasting AI Progress: Evidence from a Survey of Machine Learning Researchers

Baobao Zhang, Noemi Dreksler, Markus Anderljung, Lauren E. Kahn, Charlie Giattino, Allan Dafoe, Michael C. Horowitz

Abstract

Advances in artificial intelligence (AI) are shaping modern life, from transportation and healthcare to science and national defense. Forecasts of AI development are essential for improving policy- and decision-making. This paper reports results from a large survey of AI and machine learning (ML) researchers conducted in late 2019, eliciting forecasts for near-term milestones and human-level machine intelligence (HLMI). In aggregate, researchers placed a 50% likelihood on HLMI being achieved by 2060. While long-term timelines have remained stable since a similar 2016 study, forecasts for several near-term milestones have accelerated. Furthermore, researchers expressed significant optimism regarding the societal impact of HLMI.

Introduction

Forecasting the trajectory of AI is notoriously difficult but essential for effective policy, safety research, and economic planning. While quantitative trends like compute scaling provide some data, the subjective beliefs of the researchers building the technology offer a "wisdom of the crowds" perspective on when transformative milestones might be reached. This study updates and validates the findings of a highly-cited 2016 survey by Grace et al. (2018), investigating whether recent breakthroughs have accelerated expert timelines and how individual beliefs evolve over time.

Methodology

The researchers fielded a large-scale survey in late 2019, targeting authors who presented at two premier AI conferences: ICML and NeurIPS. The study utilized two primary groups:

  • Cross-sectional Sample: 296 researchers providing new forecasts.
  • Panel Sample: 49 individuals who participated in the 2016 survey, allowing for longitudinal comparison.

Respondents provided three points on a probability distribution (10%, 50%, and 90% likelihood) for various milestones. These were modeled using a "median parameters" method to fit a gamma Cumulative Distribution Function (CDF). The survey also included framing experiments to test if asking for "fixed years" versus "fixed probabilities" influenced the results.

Key Findings

Stability of Long-Term Forecasts

Despite significant progress in fields like natural language processing and computer vision between 2016 and 2019, the aggregate 50% probability date for HLMI remained remarkably stable, shifting only slightly from 2058 to 2060. This suggests that expert expectations for the ultimate arrival of general intelligence are not easily swayed by short-term technological hype.

Comparison of HLMI probability forecasts between the 2016 and 2019 surveys, showing high consistency in long-term expectations.

Acceleration of Near-Term Milestones

While long-term dates stayed consistent, researchers became significantly more optimistic about specific sub-tasks. Out of 18 milestones tracked, 13 saw their predicted dates move earlier. For instance, the predicted date for an AI writing a New York Times bestseller moved forward by 25 years (from 2059 to 2034), and the timeline for solving the Putnam Math Competition was reduced by 29 years. This indicates that experts see the path to HLMI becoming "denser" with intermediate achievements.

Comparison of predicted years for specific AI milestones between 2016 and 2019, illustrating a general trend toward earlier achievement dates.

Societal Impact and Cognitive Biases

Contrary to common "doomer" narratives, the median probability assigned to an "extremely bad" outcome (such as human extinction) was only 2%. The vast majority of researchers expect the impact of HLMI to be "on balance good" or "extremely good."

However, the study also revealed significant "framing effects." When asked for probabilities at fixed years, researchers gave longer timelines than when asked for years at fixed probabilities. This suggests that even experts are susceptible to cognitive biases in how they conceptualize the future.

Conclusion and Limitations

The 2019 survey confirms a median expectation for HLMI by 2060, with a notable acceleration in near-term milestones. While the cross-sectional sample is robust, the longitudinal panel remains small (n=49), limiting definitive conclusions about individual belief shifts. Future research should focus on larger longitudinal samples and investigate the psychological drivers behind framing effects to improve the calibration of expert forecasts.

Summary updated

Abstract

Advances in artificial intelligence (AI) are shaping modern life, from transportation, health care, science, finance, to national defense. Forecasts of AI development could help improve policy- and decision-making. We report the results from a large survey of AI and machine learning (ML) researchers on their beliefs about progress in AI. The survey, fielded in late 2019, elicited forecasts for near-term AI development milestones and high- or human-level machine intelligence, defined as when machines are able to accomplish every or almost every task humans are able to do currently. As part of this study, we re-contacted respondents from a highly-cited study by Grace et al. (2018), in which AI/ML researchers gave forecasts about high-level machine intelligence and near-term milestones in AI development. Results from our 2019 survey show that, in aggregate, AI/ML researchers surveyed placed a 50% likelihood of human-level machine intelligence being achieved by 2060. The results show researchers newly contacted in 2019 expressed similar beliefs about the progress of advanced AI as respondents in the Grace et al. (2018) survey. For the recontacted participants from the Grace et al. (2018) study, the aggregate forecast for a 50% likelihood of high-level machine intelligence shifted from 2062 to 2076, although this change is not statistically significant, likely due to the small size of our panel sample. Forecasts of several near-term AI milestones have reduced in time, suggesting more optimism about AI progress. Finally, AI/ML researchers also exhibited significant optimism about how human-level machine intelligence will impact society.
BibTeX
@misc{zhang2022forecasting,
  author = {Baobao Zhang and Noemi Dreksler and Markus Anderljung and Lauren E. Kahn and Charlie Giattino and Allan Dafoe and Michael C. Horowitz},
  title = {{Forecasting AI Progress: Evidence from a Survey of Machine Learning Researchers}},
  year = {2022},
  eprint = {2206.04132},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.04132},
}
Figure