This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.

Artificial Intelligence and Jobs: Evidence from Online Vacancies

Acemoglu, Autor, Hazell, Restrepo

2020NBER Working Paper Series117 citations

Observational labor marketTheoretical model

AI ExposureAI AdoptionSkills & TasksFinance

Abstract

We study the impact of AI on labor markets using establishment-level data on vacancies with detailed occupation and skill information comprising the near-universe of online vacancies in the US from 2010 onwards.There is rapid growth in AI related vacancies over 2010-2018 that is greater in AI-exposed establishments.AI-exposed establishments are reducing hiring in non-AI positions.We find no discernible relationship between AI exposure and employment or wage growth at the occupation or industry level, however, implying that AI is currently substituting for humans in a subset of tasks but it is not yet having detectable aggregate labor market consequences.

Summary

Acemoglu, Autor, Hazell, and Restrepo use establishment-level Burning Glass vacancy data and occupation-based AI exposure measures to study AI adoption patterns, finding that establishments with AI-suitable task structures increase AI hiring and reduce non-AI hiring, with significant skill changes but no detectable aggregate labor market effects.

Main Finding

A one standard deviation increase in establishment AI exposure is associated with 9-16% more AI vacancy postings, significant skill churn, and 5-12% fewer non-AI vacancies, but no detectable aggregate employment or wage effects at industry or occupation levels, suggesting AI is substituting for tasks within establishments without yet generating economy-wide impacts.

Primary Datasets

Lightcast (formerly Burning Glass)

Labor outcomes

Burning Glass (2010-2018)

Secondary Datasets

Compustat

Other

O*NET

Other

OES/OEWS

Labor outcomes

O*NET, Compustat, BLS OES

Key Methods: Establishment-level panel regressions with firm, industry, and commuting zone fixed effects; shift-share design using baseline occupation structure to construct AI exposure
Sample Period: 2010-2018
Geographic Coverage: US
Sample Size: 1.075-1.16 million establishments (depending on exposure measure) posting vacancies in 2010-12 or 2016-18; approximately 340,000 establishments posting in both periods for skill analysis
Level of Analysis: Firm, Occupation, Industry, Region
Occupation Classification: SOC 2010
Industry Classification: NAICS 2012
Replication Package: Yes

Notes

Establishment-level AI adoption measures [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks. [Claude classification]: Paper uses shift-share design with baseline occupation structure as exposure measure. Authors acknowledge this is not causal identification - it controls for observables but does not claim exogenous variation. The paper explicitly excludes AI-producing sectors (NAICS 51 and 54) to focus on AI-using establishments. Three different AI exposure indices yield varying results, with Felten et al. measure most robust. Standard errors account for shift-share structure using Borusyak et al. (2021) method in robustness checks.