Forecasting extreme labor displacement: A survey of AI practitioners

• Published in: Technological forecasting & social change Vol. 161; p. 120323
• Main Authors: Gruetzemacher, Ross, Paradice, David, Lee, Kang Bok
• Format: Journal Article
• Language: English
• Published: New York Elsevier Inc 01.12.2020; Elsevier B.V; Elsevier Science Ltd
• Subjects: AI labor displacement; Artificial intelligence; Automation; Business schools; Displacement; Forecasting; Labor; Surveys; AI; Artificial intelligence; AI labor displacement
• ISSN: 0040-1625; 1873-5509
• DOI: 10.1016/j.techfore.2020.120323

•AI practitioners give a 1-in-10 chance of developing AI systems that outperform humans at 90% of human tasks this decade.•Human-level AI practitioners make shorter and more precise labor displacement forecasts than mainstream AI practitioners.•Human-level AI practitioners give a 15-year median forecast for a 50% probability of 90% labor displacement from AI.•Extreme labor-displacing AI scenarios that fall short of human-level AI are probable enough that they should not be ignored.•Forecasts of the impact of computational resources suggest extreme labor-displacing AI to be plausible within the decade. While labor-displacing AI has the potential to transform critical aspects of society in the near future, previous work has ignored the possibility of the extreme labor displacement scenarios that could result. To explore this we surveyed attendees of three AI conferences in 2018 about near-to-mid-term AI labor displacement as well as five more extreme labor-displacing AI scenarios. Practitioners indicated that a median of 22% of tasks which humans are currently paid to do could be automated with existing AI; they anticipate this figure rising to 40% in 5 years and 60% in 10 years. Median forecasts indicated a 50% probability of AI systems being capable of automating 90% of human tasks in 25 years and 99% of human tasks in 50 years. Practitioners surveyed at the different conferences had similar forecasts for AI labor displacement this decade, but attendees of the Human-level AI Conference had significantly shorter and more precise forecasts for the more extreme labor-displacing AI scenarios. Interestingly, median forecasts of a 10% probability of 90% and 99% of human tasks being automated were 10 years and 15 years, respectively. We conclude that future of work researchers should more carefully consider these relatively high likelihoods of extreme labor-displacing AI scenarios.