A pooled review of 19 studies finds no uniform labour-market effect from AI: average impacts on employment, wages and skills are statistically indistinguishable from zero, while outcomes vary widely by sector, occupation and institutional setting.

Main Finding

The meta-analysis of 321 estimates from 19 empirical studies finds that the pooled effect of AI / automation exposure on labour-market outcomes (employment, wages, and skill demand) is small and statistically insignificant on average. However, there is substantial heterogeneity across studies: some report negative employment effects, others positive outcomes (productivity gains, wage increases, or skill upgrading). Moderator tests find no systematic difference between AI‑specific measures and broader automation indicators. Overall: effects are highly context dependent rather than uniformly displacement‑ or upgrade‑driven.

Key Points

• Data pooled: 321 effect estimates drawn from 19 empirical studies (published 2015–2025).
• Outcome coverage: employment, wages, and skill demand (separate pooled analyses for each).
• Effect-size metric: all regression coefficients converted to Partial Correlation Coefficients (PCCs) for comparability (range −1 to 1).
• Meta-analytic model: three‑level random‑effects meta‑analysis (REML) to account for sampling variance (level 1), within‑study variance among multiple estimates (level 2), and between‑study variance (level 3).
• Inference robustness: cluster‑robust standard errors (CR2) with Bell‑McCaffrey dof correction to mitigate small‑cluster bias.
• Core result: average PCC ≈ 0 (small, not statistically significant) for aggregate and outcome‑specific analyses.
• Heterogeneity: large and meaningful heterogeneity across estimates — sign and magnitude vary by sector, occupation, and institutional context.
• Moderator analysis: no consistent difference between AI‑specific exposure measures and broader automation/robotics measures.
• Contributions claimed: harmonisation of AI vs broader automation within a single PCC framework; separate pooled effects for employment/wages/skills; use of three‑level model to retain multiple estimates per study without overstating precision.

Data & Methods

• Search & selection:
  • Databases: ScienceDirect, Web of Science, EBSCOhost, Google Scholar plus grey literature and snowballing.
  • Time window: 2015–2025; English; empirical research articles.
  • Initial hits ≈ 1,678; after screening and exclusions (e.g., no regression, not comparable, missing necessary reporting) 19 studies included.
• Extracted information: regression coefficients measuring technology exposure → labour outcomes, standard errors, sample sizes, number of covariates/fixed effects (to compute degrees of freedom).
• Effect-size computation:
  • Converted regression outputs to Partial Correlation Coefficients (PCC) using t-statistics and degrees of freedom; corresponding PCC standard errors computed analytically.
  • Excluded estimates if degrees of freedom could not be reliably determined.
• Meta-analysis:
  • Three‑level random‑effects model (levels: sampling variance, within‑study variance, between‑study variance); REML estimation.
  • Cluster‑robust (CR2) SEs and Bell‑McCaffrey correction for small numbers of clusters.
  • Meta‑regressions/subgroup analyses to test differences by outcome (employment/wage/skill) and by technology domain (AI‑specific vs broader automation).
• Limitations acknowledged by authors:
  • Relatively small number of studies (19) — increases uncertainty and reduces power.
  • Heterogeneous definitions/measures of technological exposure and labour outcomes.
  • Exclusion of non‑English and pre‑2015 studies; some estimates excluded because of missing df/ reporting.
  • Potential instability of meta-analytic estimates when study count is low.

Implications for AI Economics

• No uniform macro conclusion: Policymakers and firms should not assume a single, predictable labour effect from AI adoption (neither guaranteed large displacement nor guaranteed upskilling). Effects depend on sector, occupation, firm size, institutional settings, and the specific technology/task mix.
• Research practice:
  • Future empirical work should report statistics needed for meta‑analysis (t‑stats, df, full regression details) to improve syntheses.
  • Distinguish types of AI/automation and the tasks they affect (routine vs non‑routine; cognitive vs manual) and measure mechanisms (productivity, task reallocation, complementarity).
  • Use designs that better identify causal channels (panel methods, plausibly exogenous adoption variation, firm‑level administrative data).
• Policy and management:
  • Targeted, context‑sensitive interventions (reskilling/upskilling, sectoral transition support) are more appropriate than one‑size‑fits‑all policies.
  • Monitor occupation‑ and task‑level labour demand indicators (vacancy data, detailed skill tags) to detect heterogeneous labour effects early.
  • Combine automation investments with human capital strategies to capture complementarities where they exist.
• Meta‑methodology:
  • Three‑level meta‑analytic designs and PCC harmonisation are useful standards for future syntheses in this literature.
  • Given observed heterogeneity, future meta‑analyses should prioritize moderator data collection (industry, occupational mix, institutional context, technology type, firm size) to unpack conditional effects.
• Research agenda:
  • Expand the evidence base (more studies, broader geographies, longer time horizons) to improve power and identify systematic patterns.
  • Explore dynamic and distributional consequences (longer‑run employment adjustment, within‑occupation wage dispersion, inequality effects).
  • Better measurement of AI intensity/type (task‑level exposure, model complexity, on‑the‑job tool usage) to separate AI‑specific mechanisms from general automation.

Overall takeaway: AI and automation do not produce a single, replicable labour‑market effect across contexts. Empirical and policy work should shift from searching for a universal average effect to mapping when, where, and why specific labour outcomes arise.