Artificial Intelligence and Employment: New Cross-Country Evidence

The rapid progress of AI in domains such as image and speech recognition, natural language processing, translation, and predictive analytics has renewed concerns about worker displacement and broader labor market effects. Earlier technological waves largely automated routine tasks, disproportionately affecting low- and middle-skill jobs, whereas AI can increasingly automate non-routine cognitive tasks, potentially influencing higher-skilled occupations as well. Theory suggests ambiguous employment effects: substitution may reduce employment, while productivity gains can increase demand when output expands. This study investigates the relationship between AI exposure and labor market outcomes across 23 OECD countries, focusing on employment levels and working hours. It adapts a task-based AI exposure indicator to allow cross-country and within-occupation variation, and examines heterogeneity by computer use (as a proxy for digital skills), exploring whether digital-intense occupations benefit more from AI-driven productivity gains while digitally less-intensive ones face reductions in hours.

Prior work distinguishes routine versus non-routine tasks, with earlier automation waves primarily affecting routine work (Autor, Levy, and Murnane, 2003). Robots substituted routine manual tasks (Acemoglu and Restrepo, 2020; Raj and Seamans, 2019). Recent AI advances extend potential automation to non-routine cognitive tasks (Lane and Saint-Martin, 2021). US evidence to date largely finds limited displacement from AI: Felten et al. (2019) link AI exposure to higher wage growth but not employment declines, particularly in software-intensive and high-wage occupations. Fossen and Sorgner (2019) report higher job stability and wage growth for AI-exposed individuals, especially among highly educated and experienced workers. Acemoglu et al. (2020) find AI-exposed firms adjust skill demand and restrict hiring in non-AI roles, though no clear occupational-level employment effects appear yet, possibly due to early diffusion stages. Indicators of AI deployment span demand-based measures (online job postings requiring AI skills) and task-based exposure measures (linking AI capabilities to job abilities/tasks). Demand-based proxies may miss adoption where AI development is outsourced or user interaction requires no AI skills. Task-based measures capture potential automatability but not actual adoption heterogeneity across time and place. Both approaches have blind spots, particularly where AI augments other technologies (e.g., robotics).

Design: Task-based cross-country analysis at the occupation-by-country (ISCO-08 2-digit) level for 23 OECD countries over 2012–2019. The core measure of AI exposure adapts Felten et al. (2018, 2019) by linking progress in nine AI applications (EFF) to ONET abilities, then mapping abilities to PIAAC task frequencies to construct occupation–country exposure scores scaled 0–1. Authors manually linked 35 ONET abilities (out of 52; physical/psychomotor/sensory abilities largely absent in PIAAC) to 33 PIAAC tasks, grouped into 12 categories, generating country-specific ability-use intensities by occupation. Exposure reflects where AI progress aligns with abilities required in each occupation and country. Data: Employment levels and hours from EU-LFS, ENOE (Mexico), and US-CPS; usual weekly hours are individuals’ reported typical hours; Mexico excluded from hours due to data gaps. Demand for AI-related technical skills (United Kingdom and United States only) from Burning Glass Technologies (BGT), defining AI-skill postings via keyword lists (Acemoglu et al., 2020). Controls: Offshorability (Autor and Dorn, 2013; based on Firpo et al., 2011), exposure to software and robots (Webb, 2020 via patent–task text overlap), and share of employment in tradable sectors (agriculture, industry, finance/insurance). Classification by computer use (PIAAC variable “use of a computer at work”), into low/medium/high terciles per country–occupation cells; alternative subgrouping by wage level (Goos et al., 2014) and by prevalence of creative/social tasks (derived from PIAAC frequencies). Empirical model: Cross-sectional regressions of percentage change 2012–2019 in (a) employment, (b) average usual weekly hours, and (c) share of part-time workers on 2012 AI exposure with country fixed effects and controls: Y_ij = α_j + β AI_ij + γ X_ij + u_ij, where i indexes ISCO-2 occupations and j countries. Robustness: Alternative AI exposure measures using O*NET ability prevalence/importance (US-based) instead of PIAAC mapping; additional task-based indicators (Webb, 2020; Tolan et al., 2021); weighting by inverse number of country observations; analysis by subgroups (computer use, wage level, creative/social task intensity); checks using involuntary part-time shares.

• Overall relationship: Across all occupations (2012–2019), no robust aggregate relationship between AI exposure and employment growth once controls are included.
• High computer-use occupations: Strong positive association between AI exposure and employment growth. In parsimonious models for the high computer-use subsample, the coefficient on AI exposure is about 85.7; with controls and 1-digit occupation fixed effects, coefficients remain positive and significant (e.g., 94.4 to 144.6). A one-standard-deviation increase in AI exposure among high-computer-use occupations (SD ≈ 0.067) corresponds to approximately +5.7 percentage points higher employment growth.
• Low computer-use occupations: AI exposure is associated with larger declines in average weekly hours. In low-computer-use subsample, coefficients around −4.8 imply that a one-standard-deviation increase in exposure (SD ≈ 0.125) is linked to about −0.60 percentage points greater decline in usual weekly hours (≈13 minutes per week at ~37.2 hours average). The increase in part-time employment shares is consistent with this pattern (significant positive coefficients in low computer-use subsample), including for involuntary part-time, though some specifications lose significance when adding all controls and occupation fixed effects.
• Demand for AI-related technical skills: Positive but modest correlation between AI exposure and growth in AI-skill job postings, especially in high-computer-use occupations. However, magnitudes are small relative to overall employment changes. In 2019, AI-skill postings comprised on average ~0.24% of US postings (0.10% in 2012) at the occupation level; at the sector level, UK rose from ~0.14% (2012) to ~0.67% (2019) and US from ~0.26% to ~0.94% (2019). Thus, increased AI-skill demand cannot explain the full employment gains observed.
• Sector-level consistency: Across sectors in the UK and US, higher task-based AI exposure in 2012 is associated with larger 2012–2019 increases in AI-skill posting shares (e.g., in the UK, a 1 SD increase in exposure ≈ +0.33 percentage point change in AI-skill share), suggesting consistency between task-based exposure and demand-based indicators at sector level.
• Cross-sectional patterns: AI exposure is higher in highly educated, white-collar occupations (e.g., business and science/engineering professionals, managers) and lower in physically intensive occupations (e.g., cleaners, laborers). Cross-country differences in exposure are modest relative to cross-occupation differences; Northern European countries show somewhat higher average exposure than Eastern European ones.
• Descriptives: Employment grew by ~10.8% on average across occupations and countries (2012–2019), while average usual weekly hours fell by ~0.40% (~9 minutes/week).

Findings suggest heterogeneous employment effects of AI depending on digital intensity. In occupations with high computer use (a proxy for stronger digital skills and greater familiarity with software-like tools), AI exposure is linked to higher employment growth, consistent with productivity-enhancing effects outweighing substitution. Partial automation may shift task composition toward higher value-added, non-automatable tasks (e.g., complex analysis, creativity, social interaction), raising overall labor demand when product demand is elastic. In contrast, in low computer-use occupations, AI exposure correlates with reduced working hours and increased part-time incidence, implying that substitution and limited complementarities may dominate when workers lack the digital skills to leverage AI effectively. The positive but small increase in AI-skill postings relative to large employment changes indicates that most gains are not driven by direct hiring into AI-technical roles; instead, complementarities between AI tools and digitally proficient workers in non-AI roles likely underpin the observed patterns. These results align with prior US evidence (e.g., wage growth and stability in AI-exposed jobs) and reinforce the role of skill-biased technology adoption. However, estimates are associative, not causal, and sensitive to measurement of exposure and controls in some specifications. Policy implications include strengthening digital skills and enabling task reallocation within occupations to facilitate AI complementarity, especially for mid- and low-digital-intensity roles.

The paper extends a task-based AI exposure measure to 23 OECD countries and links it to employment and hours outcomes from 2012 to 2019. Overall, there is no clear aggregate relationship between AI exposure and employment growth. Yet, in high computer-use occupations, greater AI exposure is associated with higher employment growth, whereas in low computer-use occupations, exposure relates to declines in average working hours and more part-time work. These patterns are consistent with AI-driven partial automation that boosts productivity and shifts tasks toward higher value-added activities for digitally proficient workers, while workers with weaker digital skills may face reduced hours. Future research should: (1) identify causal mechanisms and dynamics as AI adoption diffuses; (2) integrate direct adoption measures with task-based exposure; (3) better capture AI’s augmentation of other technologies (e.g., robotics) and physical abilities; (4) analyze distributional impacts across worker groups and countries; and (5) evaluate policies that enhance digital skills and task reallocation within occupations.

• Exposure vs. adoption: Task-based measures capture potential automatability, not actual AI adoption or timing, limiting inference on real-time deployment across countries and sectors.
• Ability mapping constraints: The PIAAC-based linkage omits 17 physical/psychomotor/sensory abilities due to data unavailability, potentially understating exposure in physically intensive occupations and missing AI’s augmentation of robotics.
• External controls’ transferability: Measures for offshorability, software, and robot exposure are derived from US O*NET/patent data and applied cross-country, assuming similar cross-occupation distributions.
• Demand-based data limits: BGT postings reflect vacancies, not hires; coverage biases exist across sectors and occupations; pre-2018 data limited mainly to English-speaking countries; occupation-level AI adoption may not require AI-skill postings.
• Outcome measurement: Hours analysis excludes Mexico; occupation coding at ISCO-2 aggregates diverse roles, possibly diluting within-occupation variation; measurement error could attenuate coefficients, especially with occupation fixed effects.
• Causality: Cross-sectional associations with country fixed effects do not establish causal effects; some findings (e.g., hours/part-time) lose significance with extensive controls or alternative exposure indicators.
• Scope and period: Analysis is limited to 2012–2019; AI diffusion post-2019 (e.g., rapid generative AI advances) is not captured.