Source-linked AI summary
When Will AI Exceed Human Performance? Evidence from AI Experts
Katja Grace, John Salvatier, Allan Dafoe, Baobao Zhang, Owain Evans
TL;DR
The paper addresses the need for better forecasts of transformative AI and surveys machine-learning researchers about AI timelines and consequences. Using responses from researchers at major machine-learning conferences, it estimates timelines for capabilities, HLMI, and labor automation and elicits views about explosive progress and outcomes. Respondents placed a 50% probability on HLMI within 45 years and full labor automation 122 years from now, while judging explosive progress possible but improbable and catastrophic outcomes possible.
Problem
Accurate forecasts of transformative AI are needed to anticipate its social, economic, legal, and policy consequences.
Method
The paper surveys researchers who published at NIPS or ICML, eliciting probability distributions over AI timelines and judgments about advanced AI’s social and ethical impacts.
Results
Respondents assigned a 50% probability to HLMI within 45 years and full labor automation 122 years from now, while viewing explosive progress after HLMI as possible but improbable.
Takeaways & Limitations
The findings provide expert forecasts for discussions among researchers and policymakers about anticipating and managing AI trends, including the need to consider both benefits and catastrophic risks.
Takeaways & Limitations
The sample represents machine learning more directly than artificial intelligence as a whole, and unmeasured non-response bias cannot be ruled out.
Abstract
from arXiv · showhide
Advances in artificial intelligence (AI) will transform modern life by reshaping transportation, health, science, finance, and the military. To adapt public policy, we need to better anticipate these advances. Here we report the results from a large survey of machine learning researchers on their beliefs about progress in AI. Researchers predict AI will outperform humans in many activities in the next ten years, such as translating languages (by 2024), writing high-school essays (by 2026), driving a truck (by 2027), working in retail (by 2031), writing a bestselling book (by 2049), and working as a surgeon (by 2053). Researchers believe there is a 50% chance of AI outperforming humans in all tasks in 45 years and of automating all human jobs in 120 years, with Asian respondents expecting these dates much sooner than North Americans. These results will inform discussion amongst researchers and policymakers about anticipating and managing trends in AI.
Introduction
The paper surveys machine-learning researchers about AI progress, its timing, and its potential consequences to improve forecasting for policy and societal preparation. Respondents forecast rapid advances in specific capabilities, while assigning later timelines to HLMI and full labor automation, with substantial regional variation.
- Motivation: AI advances could create major employment, infrastructure, cybersecurity, legal, military, and marketing challenges, making accurate forecasting valuable for preparation.The paper frames forecasting as relevant to both policymakers and AI developers.
- Survey scope: The survey sampled researchers who published at NIPS or ICML and received 352 responses, covering AI capabilities, occupations, automation, and social impacts.These conferences are described as premier venues for peer-reviewed machine-learning research.
- Forecasts: 50% of respondents’ aggregate probability placed HLMI within 45 years, while full labor automation received a 50% probability 122 years from now.HLMI was defined as unaided machines accomplishing every task better and more cheaply than human workers; full labor automation used an equivalent occupational criterion.
- Intelligence explosion: Researchers assigned a 10% median probability to AI becoming vastly better than humans in all tasks two years after HLMI, versus 20% for explosive global technological improvement.Both estimates were considered possible but improbable, with interquartile ranges of 1–25% and 5–50%, respectively.
- Outcomes and safety: Median probabilities were 25% for a good long-run outcome, 20% for an extremely good outcome, 10% for a bad outcome, and 5% for an extremely bad outcome such as human extinction.The survey used a five-point scale for possible long-run effects of HLMI on humanity.
- Variation and implications: Asian respondents expected HLMI in 30 years, compared with 74 years for North Americans, while 48% favored increasing societal priority for AI-risk research.The regional comparison used undergraduate institution as a proxy for country of origin, and the authors note possible unmeasured non-response bias.
Supplementary Information
The survey covered AI timelines, intelligence-explosion probabilities, welfare implications, research inputs, and narrow AI milestones.
- HLMI questions used three framings: direct HLMI timing, automation of all human occupations, and extrapolation from recent AI progress.
- Researchers were asked about the probability of an “intelligence explosion.”
- One question assessed the welfare implications of HLMI.
- The survey examined how inputs such as hardware progress affect the rate of AI research.
5. Two questions about sources of disagreement about AI timelines and “AI Safety.”
This section defines HLMI, describes how predictions were elicited and aggregated, and records additional survey coverage of AI milestones, safety, and demographics.
- Survey Content: The survey also included 32 narrow AI milestone questions, AI Safety questions, and demographic questions administered online.
- HLMI means a machine that is better than humans at all tasks.
- Respondent Data: Citation count and seniority were collected from public sources, with seniority measured by years since PhD start.
- Elicitation of Beliefs: Timeline questions used fixed-probability and fixed-years framings, which sampled different CDF points and produced framing effects among respondents.
- Analysis: Aggregate distributions fitted a gamma CDF to each respondent, combined individual fits into a mixture distribution, and used bootstrapping for 95% confidence intervals.
Supplementary Figures
The supplementary figures show aggregate HLMI forecasts by demographic group and compare two question framings with their averaged forecast.
- Figure S1 groups aggregate HLMI forecast CDFs by country, seniority quartile, and citation-count quartile.
- Figure S2 compares aggregate forecasts from fixed-probability and fixed-years framings and displays their combined average.
Supplementary Tables
The supplementary material provides implementation details, respondent comparisons, survey-response summaries, and operational descriptions of AI milestones.
- Automation Predictions: The NA/Asia gap subtracts Asian median estimates from North American median estimates for job-automation timelines.
- Automation Predictions: Table S1 reports median years from 2016 for human-job automation by undergraduate-region grouping.
- HLMI Regression: Robust regression models individual HLMI timelines using gender, citations, seniority, question framing, and undergraduate region.
- Respondent Comparison: The respondent comparison includes 406 respondents and 399 randomly sampled non-respondents from NIPS/ICML authors.
- AI Milestones: The milestone descriptions define capabilities spanning novel-category learning, one-shot learning, video generation, speech transcription, and text-to-speech.
- AI Milestones: Additional milestones cover mathematical research, Putnam-level mathematics, Go, StarCraft, random-game play, Atari, laundry folding, city racing, LEGO assembly, sorting, and code generation.