[Paper Review] AI exposure predicts unemployment risk

Morgan R. Frank, Yong‐Yeol Ahn+1 more|arXiv (Cornell University)|Aug 4, 2023

Digital Economy and Work Transformation4 citations

TL;DR

This study evaluates multiple AI exposure scores to predict unemployment risk, job separations, and within-occupation skill changes using high-resolution, state-level unemployment insurance data from 2010–2020. While individual AI exposure models show no predictive power, an ensemble of these models significantly predicts unemployment risk, highlighting the need for dynamic, context-aware, and validated AI exposure assessment methods across occupations and regions.

ABSTRACT

Is artificial intelligence (AI) disrupting jobs and creating unemployment? Despite many attempts to quantify occupations' exposure to AI, inconsistent validation obfuscates the relative benefits of each approach. A lack of disaggregated labor outcome data, including unemployment data, further exacerbates the issue. Here, we assess which models of AI exposure predict job separations and unemployment risk using new occupation-level unemployment data by occupation from each US state's unemployment insurance office spanning 2010 through 2020. Although these AI exposure scores have been used by governments and industry, we find that individual AI exposure models are not predictive of unemployment rates, unemployment risk, or job separation rates. However, an ensemble of those models exhibits substantial predictive power suggesting that competing models may capture different aspects of AI exposure that collectively account for AI's variable impact across occupations, regions, and time. Our results also call for dynamic, context-aware, and validated methods for assessing AI exposure. Interactive visualizations for this study are available at https://sites.pitt.edu/~mrfrank/uiRiskDemo/.

Motivation & Objective

• To assess whether existing AI exposure scores predict unemployment risk, job separations, and skill changes at the occupation level.
• To overcome limitations of prior studies that rely on aggregate employment or wage data by using high-resolution, state-level unemployment insurance data from 2010 to 2020.
• To evaluate the predictive validity of multiple AI exposure models, including those from Acemoglu & Autor, Frey & Osborne, Arntz et al., Brynjolfsson et al., Felten et al., and Webb, across different labor outcomes.
• To identify geographic and occupational heterogeneity in AI exposure model applicability, challenging the use of single-score assessments.
• To advocate for dynamic, context-aware, and empirically validated methods for measuring AI exposure in labor market policy and research.

Proposed method

• Constructed a high-resolution dataset of monthly unemployment counts by most-recent occupation from each U.S. state’s unemployment insurance office (2010–2020).
• Collected and standardized multiple AI exposure scores from established studies, including computer usage, routine task exposure, automation probability, suitability for machine learning, and AI/software exposure percentages.
• Applied multiple linear regression models with fixed effects for year and major SOC categories to isolate the predictive power of each AI exposure score on unemployment risk, job separations, and within-occupation skill change.
• Used standardized, centered variables to ensure comparability across different AI exposure metrics with varying units and scales.
• Evaluated model performance using R² and adjusted R², with significance levels reported for each coefficient.
• Conducted ensemble modeling by combining individual AI exposure scores to assess collective predictive power.

Figure 1: Technological exposure scores are not consistent with each other and cannot individually predict unemployment risk well. (A) The Pearson correlation of pairs of AI exposure scores across occupations. Scores are ordered according to the study’s Wave (see Table 1 ). (B) The variation in unem

Experimental results

Research questions

• RQ1 Do individual AI exposure scores predict increased unemployment risk at the occupation level?
• RQ2 Can AI exposure scores predict job separation rates or within-occupation skill changes?
• RQ3 Does combining multiple AI exposure scores through an ensemble model improve predictive accuracy for unemployment risk?
• RQ4 How does the predictive power of AI exposure models vary across different U.S. states and occupations?
• RQ5 To what extent do employment and wage data obscure other labor market disruptions caused by AI, such as job separations and unemployment?

Key findings

• Individual AI exposure models, including those from Acemoglu & Autor, Frey & Osborne, Arntz et al., Brynjolfsson et al., Felten et al., and Webb, showed no significant predictive power for unemployment risk, job separations, or skill change when used in isolation.
• An ensemble model combining all AI exposure scores significantly predicted unemployment risk, with an adjusted R² of 0.304 in the final model, indicating strong collective explanatory power.
• The ensemble model remained predictive even after controlling for regional fixed effects, temporal trends, and occupational skill requirements, suggesting that multiple models capture complementary aspects of AI exposure.
• There was strong geographical heterogeneity in model applicability, with different AI exposure scores being more relevant in different U.S. states, indicating that a one-size-fits-all approach is inadequate.
• The study found that employment and wage data alone may miss critical labor market disruptions such as job separations and unemployment, which are better captured by unemployment insurance data.
• Within-occupation skill change, measured via O*NET data from 2010 to 2017, showed significant variation, and several AI exposure scores (e.g., Brynjolfsson et al.’s suitability for ML and Felten et al.’s AI exposure) were significantly associated with skill change, with coefficients of 0.330*** and -0.269***, respectively.

Figure 2: Combining AI exposure scores from all studies substantially improves predictions of unemployment risk. (A) A schematic for the theoretical influence among education, skills, technology, and unemployment risk. (B) Combined into a single linear model, AI exposure scores capture 29.1% of the