[Paper Review] Forecasting AI Progress: Evidence from a Survey of Machine Learning Researchers

Baobao Zhang, Noemi Dreksler+5 more|arXiv (Cornell University)|Jun 8, 2022

Big Data and Business Intelligence22 citations

TL;DR

The paper reports a 2019 survey of AI/ML researchers forecasting HLMI around 2060, compares with 2016 Grace et al., finds forecast stability for HLMI and earlier near-term milestones, and assesses perceived societal impact and framing effects.

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.

Motivation & Objective

• Assess trajectories of AI progress through forecasts from AI/ML researchers at ICML/NeurIPS.
• Compare 2019 cross-sectional forecasts to 2016 Grace et al. results and to recontacted panel responses.
• Evaluate researchers' views on the long-run societal impact of human-level machine intelligence.
• Examine how framing and respondent characteristics relate to forecast timelines.
• Inform governance, policy, and research priorities by understanding expert foresight.

Proposed method

• Elicited forecasts from 296 cross-sectional authors and 49 recontacted panel respondents using gamma-distribution CDF aggregation (median parameters method).
• Constructed a continuous probability distribution (CDF) from three discrete forecasts per respondent and bootstrapped 95% CIs for group medians.
• Compared cross-sectional results with Grace et al. (2018) and used LOESS as a robustness check.
• Analyzed near-term AI milestones (eighteen to twenty-two tasks) under fixed-year and fixed-probability framings.
• Applied randomization tests using Wasserstein and Kolmogorov–Smirnov statistics to compare forecast distributions across surveys.

Experimental results

Research questions

• RQ1 What is the aggregate timing for human-level machine intelligence (HLMI) to occur among AI/ML researchers?
• RQ2 How do 2019 forecasts compare with the 2016 Grace et al. results and the Grace et al. panel recontacts?
• RQ3 What is the perceived long-run impact of HLMI on humanity according to AI/ML researchers?
• RQ4 Do near-term AI milestone forecasts shift toward earlier dates, and what factors explain these changes?
• RQ5 How do framing effects (fixed years vs fixed probabilities) influence expert forecast timelines?

Key findings

• The 2019 cross-sectional sample yields a 50% probability of HLMI by 2060, with no significant difference from the 2016 Grace et al. result (2058).
• For the recontacted Grace et al. panel, the 50% HLMI forecast shifts from 2062 to 2076, but the change is not statistically significant due to small sample size.
• Respondents exhibit net positive views on HLMI’s long-run impact, with the median outlook spanning between "on balance good" and "more or less neutral."
• In 2019, many AI milestones are forecast to occur earlier than in 2016, including milestones like writing a New York Times bestselling novel (50% by 2034 vs 2059 in 2016) and other tasks such as assembling LEGO sets or composing Top-40 songs (examples noted in the text).
• Two milestones have actually been achieved since 2016: DeepMind’s Agent57 beat the Atari benchmark (publicly announced in 2020) and Facebook/CMU’s Pluribus beat top poker players (publicly announced in 2019); the 2016 forecast for Win at Atari was 50% by late 2033, while 2019 forecasts predicted 50% by 2024.
• The 2016 forecast expected 3.92 milestones to resolve vs. 2 milestones resolved by Jan 1, 2022; the 2019 forecast expected 1.65 milestones to resolve vs. 1 milestone resolved by Jan 1, 2022.