Artificial Intelligence and Employment: New Cross-Country Evidence

The study examines how recent advances in AI are linked to employment outcomes across countries. Unlike earlier automation waves that mainly affected routine tasks, current AI progress increasingly touches non-routine cognitive tasks typically found in medium- and high-skill occupations. The employment impact of AI is theoretically ambiguous: automation can substitute for workers, reducing employment, while productivity gains can increase output and labor demand if demand is sufficiently elastic. Given limited empirical evidence outside the United States, the paper’s purpose is to assess, in a cross-country setting, whether occupations more exposed to AI have experienced differential changes in employment and working hours, and whether effects differ by the intensity of computer use (a proxy for digital skills) within occupations.

Prior evidence, largely from the United States, suggests limited displacement from AI to date and some positive wage or stability effects concentrated among higher-skilled workers. Felten et al. (2019) found no aggregate employment link but positive wage growth associated with AI exposure, especially in high-wage, software-intensive occupations. Fossen and Sorgner (2019) reported higher employment stability and wages for individuals exposed to AI, with stronger effects for more educated and experienced workers. Acemoglu et al. (2020) showed that firms exposed to AI adjust hiring and skill demands but found no clear occupation-level employment effects, noting that detectable impacts may emerge as adoption spreads. The paper also reviews two indicator families to proxy AI deployment: demand-based indicators derived from online job postings (e.g., Burning Glass) and task-based indicators mapping AI capabilities to occupational task structures (e.g., Felten et al.; Brynjolfsson and Mitchell; Tolan et al.; Webb). Demand-based indicators can miss AI use when firms train in-house, outsource AI, or when interacting with AI does not require AI skills. Task-based indicators measure potential exposure rather than actual adoption and miss AI augmenting other technologies (e.g., robotics). At the sector level in the UK and US, task-based exposure correlates positively with growth in AI-skill postings, suggesting some consistency, but occupational-level demand indicators remain incomplete proxies for AI deployment.

Design: The study extends the Felten, Raj, and Seamans AI occupational exposure measure to 23 OECD countries and links it to labor market outcomes from 2012 to 2019 across 36 ISCO-08 2-digit occupations (excluding IT technology professionals and IT technicians in main analyses, with robustness including them).

AI exposure construction: Starting from Felten et al. (2018, 2019), which links progress in nine AI applications (EFF dataset, 2010–2015) to 52 ONET abilities via a crowdsourced application–ability matrix, the authors re-map abilities to country-specific task use using the OECD Survey of Adult Skills (PIAAC). They: (1) compute for each country and occupation the average frequency with which workers perform 33 PIAAC tasks (2012; exceptions: Hungary 2017, Lithuania 2014, Mexico 2017); (2) manually link PIAAC tasks to ONET abilities (binary necessity mapping), yielding 35 abilities with PIAAC coverage (17 physical/sensory abilities lack PIAAC task links); (3) aggregate within 12 task categories to avoid overweighting categories with more items; and (4) combine ability intensities with AI application progress to produce an occupation–country AI exposure score scaled 0–1. This measure captures potential automation by AI via abilities where AI has progressed, allows variation within occupations across countries, and isolates AI (not robotics) capabilities.

Data: Employment and hours outcomes come from EU-LFS, US-CPS, and Mexico’s ENOE (hours exclude Mexico due to data gaps). Average usual weekly hours are individuals’ reported usual hours. Part-time is defined as 30 hours or less per week; involuntary part-time is identified where available. Burning Glass Technologies (BGT) job postings (UK, US only) measure the share of postings requiring AI-related technical skills (keyword list per Acemoglu et al. 2020) by occupation. Additional indicators include exposure to software and to industrial robots (Webb 2020, patent–task text overlap), and offshorability (Firpo, Fortin, Lemieux; Autor and Dorn). Exposure to international trade is proxied by the occupation’s employment share in tradable sectors (agriculture, industry, finance/insurance).

Empirical strategy: For country j and occupation i, the main regression relates percentage change (2012–2019) in employment levels to 2012 AI exposure, with country fixed effects and controls: exposure to software and robots, offshorability, share employed in tradable sectors, and 1-digit ISCO dummies in some specifications. Parallel regressions use change in average usual weekly hours and change in the share of part-time workers as dependent variables. Heterogeneity analyses split occupation–country cells into terciles of computer use (from PIAAC question “Do you use a computer in your job?”). Additional splits use average wage level (Goos et al. classification), and prevalence of creative and social tasks (from PIAAC frequencies). Robustness checks use: (a) O*NET-based ability prevalence/importance (US weights) instead of PIAAC task links; (b) alternative AI exposure indicators (Webb 2020; Tolan et al. 2021); (c) alternative part-time definition (involuntary part-time).

• Descriptives: Across the 23 countries, employment grew in most occupations between 2012 and 2019; the unweighted average growth across occupations and countries was 10.8%. Average usual weekly hours fell slightly by 0.40% (about 9 minutes per week at the sample mean of 37.7 hours). AI-related skill postings increased in almost all occupations in the US and UK, but remained a tiny share of postings in 2019 (US average 0.24% vs. 0.10% in 2012; UK average 0.14% vs. 0.14% in 2012 rising to 0.67% sectorally; occupation-level averages remain below 1%).
• AI exposure patterns: AI progress (2010–2015) most strongly maps to non-routine cognitive abilities (e.g., information ordering, memorisation, perceptual speed), making highly educated white-collar occupations (business, legal, science/engineering professionals, managers) most exposed, and physical/manual occupations (e.g., cleaners, laborers) least exposed. Cross-country variation in exposure within an occupation is modest relative to cross-occupation differences.
• Employment: In pooled regressions, the simple positive correlation between AI exposure and employment growth becomes non-significant once controls are added. However, splitting by computer use shows a robust positive association among high computer-use occupations: a one standard deviation increase in AI exposure is associated with approximately 5.7 percentage points higher employment growth (coefficient about 85.7 with sd ≈ 0.067). This relationship holds with controls for trade exposure, offshorability, software/robot exposure, and broad occupation fixed effects. Relationships by wage level or by prevalence of creative/social tasks are weaker and less robust.
• Working hours and part-time: AI exposure is associated with larger declines in average usual weekly hours, particularly in low computer-use occupations. Among low computer-use occupations, a one standard deviation increase in AI exposure corresponds to an additional 0.60 percentage point drop in usual weekly hours (about 13 minutes per week at 37.2 hours), robust to controls except when adding 1-digit occupation fixed effects. The decline in hours appears driven by growth in part-time employment (including involuntary part-time), with positive and significant coefficients for low computer-use occupations in models without full fixed effects.
• AI skills demand: Growth in the share of AI-skill postings is positively (though modestly) correlated with AI exposure, especially in high computer-use occupations, but the magnitude is far too small to account for the employment growth observed in those occupations. Sector-level analyses in the UK and US also show that sectors with higher task-based AI exposure had larger increases in AI-skill postings, while some highly exposed sectors (e.g., education, energy, public administration, real estate) show little AI-skill posting growth, consistent with outsourcing or non-technical user interaction with AI. Overall interpretation: There is no clear overall link between AI exposure and employment growth across all occupations. Yet, in computer-intensive occupations, AI exposure aligns with higher employment growth, consistent with productivity/complementarity effects for digitally skilled workers. Conversely, in low computer-use occupations, AI exposure correlates with reduced working hours, suggestive of substitution effects outweighing productivity gains.

The findings align with a nuanced view of AI’s labor market impact: effects depend on occupational task content and workers’ digital capabilities. In occupations where computer use is high, workers may more easily integrate AI tools, shift toward higher value non-automatable tasks, and realize productivity gains that, together with elastic product demand, expand employment. The positive association between AI exposure and employment growth within high computer-use occupations, while modest, persists after controlling for alternative technologies (software, robots), offshorability, and tradability, suggesting a role for AI-specific complementarities. In low computer-use occupations, exposure to AI is associated with declines in working hours and increases in part-time work, indicating that AI-enabled partial automation may reduce task time without sufficient offsetting demand or within-occupation task upgrading among workers with weaker digital skills. These intensive-margin effects are consistent with substitution dominating productivity benefits in such settings. The small scale of AI-skill postings relative to total employment change implies that employment effects are not primarily driven by hiring for AI technical roles within the same occupations. Instead, employment gains likely come from broader productivity improvements and task reallocation within occupations that are complementary to general digital proficiency rather than technical AI development skills. Overall, the results suggest that digital skill readiness conditions AI’s employment effects: stronger digital environments facilitate positive employment outcomes from AI exposure, while weaker ones may experience reductions in hours. This highlights the importance of policies that improve workers’ digital skills and support task transitions within occupations.

The study extends a task-based AI exposure measure to 23 OECD countries and links it to 2012–2019 employment outcomes. Across all occupations, there is no clear overall relationship between AI exposure and employment growth. However, in high computer-use occupations, higher AI exposure is associated with stronger employment growth, while in low computer-use occupations, higher AI exposure correlates with larger declines in average working hours. A plausible mechanism is that partial AI automation raises productivity and shifts work toward higher value tasks for digitally capable workers, offsetting substitution effects; conversely, workers with weaker digital skills may not capture these benefits, leading to reduced hours and more part-time work. Future research should: (1) improve measures of actual AI adoption and usage at the occupation level (beyond task-based exposure), including AI augmenting other technologies; (2) explore causal identification strategies as AI diffusion deepens; (3) examine heterogeneity by specific digital skill profiles and training; and (4) extend analysis to additional countries and time periods as data mature.

• Measurement: The primary AI measure is task-based exposure, not actual adoption; it cannot capture uneven diffusion across time, sectors, or countries, nor AI augmenting other technologies (e.g., robotics). PIAAC lacks many physical/sensory tasks, omitting 17 O*NET abilities and potentially understating exposure in physically intensive occupations. Alternative indicators (Webb; Tolan) differ in construct (patents vs. research intensity) and yield less precise or non-significant results in some subsamples.
• Data scope: Hours analyses exclude Mexico. AI-skill postings are only available for the UK and US and represent a small share of postings; BGT data have known representativeness limitations across occupations and sectors.
• External indicators: Controls for software/robot exposure and offshorability rely on US-based occupation measures applied to all countries, potentially introducing measurement error.
• Identification: Results are associational with country fixed effects; some relationships lose significance when adding broad occupation fixed effects, suggesting limited within-group variation or attenuation bias. Country-case plots show heterogeneous patterns.
• Coverage: The sample excludes IT professionals and technicians in the main analysis (though robustness suggests results are not sensitive), and some occupation–country cells are missing due to data limitations. The period of AI progress (to 2015) and outcomes (to 2019) may still be early relative to AI diffusion dynamics.