INTRODUCTION

Artificial intelligence (AI) has been emerging as a major technological force with the potential to transform labor markets across various industries and occupations. As AI’s ability to perform routine tasks increases, concerns regarding how technological innovation may affect labor demand and employment across different economies have grown. Past empirical research on large exogenous economic shocks provides useful frameworks for studying their macroeconomic effects. For example, Autor, Dorn, and Hanson (2013) show that U.S. regions more exposed to Chinese imports experienced larger job losses and labor market disruptions. Their findings demonstrate how differences in economic structure can cause global shocks to have heterogeneous labor-market effects.
Our empirical study applies a similar framework to the global expansion of artificial intelligence since 2018, and especially to the expansion of large language models after 2022. These developments represent major technological shifts influencing labor demand across occupations and sectors. This leads to our research question: How does exposure to artificial intelligence affect unemployment across countries, and do countries with higher AI exposure experience differential labor-market outcomes following the diffusion of AI technologies after 2018 and the expansion of large language models after 2022?
To answer this question, we adopt an interaction framework inspired by Rajan and Zingales (1998) and use cross-country differences in occupational composition to measure structural exposure to AI. If AI affects labor demand, countries whose employment structures are more concentrated in AI-exposed occupations should experience greater unemployment changes.

This paper investigates the effect of AI exposure on national-aggregate unemployment rates across 133 countries from 2010 to 2024. We constructed country-level exposure indices by aggregating occupational exposure scores from Felten et al. (2021) and Eloundou et al. (2023) to the 2-digit ISCO-08 level, using employment shares from ILO data. The correlation between these two indices is very high (r = 0.99), indicating that the occupational composition of an economy, rather than the scoring methodology, largely drives country-level exposure. We estimated a 2SLS regression with country- and year fixed effects, instrumenting the interaction between a post-2022 dummy and the country-level FRS index with the Oxford Government AI Readiness Index interacted with the same dummy. The main finding is that more-exposed countries are associated with lower unemployment after 2018: −1.82 percentage points per unit increase in FRS exposure (p < 0.01). This finding is robust to alternative occupational scoring, an alternative instrumental variable (the IMF AI Preparedness Index 2023), and the exclusion of the COVID-19 years. In all six specifications, the coefficient on the post-2022 LLM period is positive but not statistically significant. We treat this null finding as indicative rather than conclusive, given that LLM emergence occurred just three years ago, and it agrees with the results of a company-level analysis (Anthropic 2026) that detects no impact until 2025. Overall, the findings suggest that, so far, AI may have affected unemployment mainly through task reallocation and augmentation rather than displacement. It is important to note that, according to the appendix, a pre-trend placebo test suggests that more-exposed countries had a distinctive unemployment trajectory before 2018. As a result, that fact can partly drive the post-2018 coefficients. This paper demonstrates that national unemployment rates cannot detect short-run effects of LLMs and emphasizes the need for future research using firm-, sector-, and task-level data.