Firms hiring for AI skills boost output and productivity, but the gains are not evenly shared: employment and pay fall persistently, especially for non-AI roles and junior staff.

Main Finding

AI adoption—measured as the share of establishment-level job postings that explicitly require AI-related skills (2017–2025, 13 U.S. industries)—raises labor productivity and real output but produces substantial labor displacement. Employment and total compensation fall persistently in the short and medium run, with losses concentrated among workers performing non-AI tasks and in non–senior-level positions. Gains from AI are therefore asymmetric across worker types.

Key Points

• Measurement: AI adoption is proxied by the share of an establishment’s job postings that require AI skills.
• Aggregate effects:
  • Positive: increases in labor productivity and real output following higher AI adoption.
  • Negative: persistent declines in employment and total compensation in the short-to-medium run.
• Distributional effects:
  • Employment and pay declines are concentrated among workers in non-AI tasks and in non-senior roles.
  • Senior-level and AI-task workers capture a disproportionate share of gains.
• Timing: effects documented using dynamic (panel) local projections over horizons up to the medium run (2017–2025).
• Net welfare implication: productivity/output gains coexist with notable labor market disruption and unequal distribution of benefits.

Data & Methods

• Data:
  • Establishment-level job-posting data across 13 U.S. industries, 2017–2025.
  • AI adoption indicator = share of postings explicitly requiring AI-related skills.
  • Outcomes: establishment-level labor productivity, real output, employment, total compensation; heterogeneity by task type (AI vs non-AI) and worker seniority.
• Empirical method:
  • Panel local projections to estimate dynamic responses of outcomes to changes in AI-adoption measure.
  • Heterogeneity analysis by task content and worker seniority to identify distributional impacts.
• Identification & limitations (inherent to this approach):
  • Local projections trace dynamic correlations but rely on exogeneity (or conditional exogeneity) of within-establishment variation in AI posting shares for causal interpretation.
  • Potential concerns: reverse causality (more productive firms both adopt AI and post AI-skilled jobs), measurement error (job postings capture demand for AI skills but may miss in-place use), and selection (which establishments post AI-skilled jobs).
  • Robustness steps likely include fixed effects, controls and pre-trend checks (details not provided here).

Implications for AI Economics

• Technology vs. labor tradeoff: AI can boost aggregate productivity and output while reducing labor demand and compensation in affected segments—consistent with task-based models of automation and skill-biased technological change.
• Distributional policy:
  • Need targeted policies to support workers displaced from non-AI tasks and junior positions (retraining, wage insurance, placement services).
  • Consider redistributive measures or labor-market supports to share gains more broadly (progressive taxes, social safety nets, wage subsidies).
• Labor-market institutions:
  • Strengthen upskilling and lifelong learning programs that lower frictions to transition into AI-complementary roles.
  • Encourage firm investments in complementary human capital so productivity gains translate into broader wage gains.
• Measurement & monitoring:
  • Job-posting–based AI measures are useful for timely monitoring but should be complemented with on-the-job usage data and matched employer-employee records to track actual impacts on earnings and career trajectories.
• Research directions:
  • Examine long-run equilibrium effects: do displaced workers transition to new occupations or face persistent scarring?
  • Investigate mechanisms: capital deepening vs task substitution, complementarities across tasks, firm entry/exit.
  • Improve causal identification (instrumental variables, natural experiments, event studies) to isolate causal effects of AI adoption.

Overall, the study highlights a central tension for policymakers: AI raises productivity and output but generates concentrated, persistent labor displacement that calls for active labor-market and redistribution policies to achieve inclusive gains.