Your Job at Risk? A Systematic Examination and Reconstruction of Occupational AI Exposure Indices through Convergence and Factor Analysis

Abstract

What occupations are most exposed to Artificial Intelligence (AI) capabilities? With a growing public concern over future occupational prospects under AI advancements, studies that focused on constructing the Occupational AI Exposure Indices have proliferated over the past few years, generating significant public interest and academic discourse. More importantly, these exposure indices have been increasingly adopted in the empirical economic literature that evaluates AI’s labor market impact. With a variety of occupational indices constructed following distinct methodologies and data sources, a key question that follows is: how much do they agree with each other’s ratings? How sensitive are empirical estimates of AI’s labor market effect to the choice of exposure indices? To answer these questions, I compiled 8 major exposure indices from the most recent wave of AI exposure studies for a systematic analysis of their degree of rating and inference convergence. I first conducted a correlation and clustering analysis, discovering mixed results in the degree of ranking consistency, with certain indices showing no significant correlation or a low level of correlation with all other indices. I then tested the estimates’ robustness to the choice of indices in three different empirical settings. Two of these settings saw a relatively prominent sensitivity to the selection of exposure indices in terms of significance or size of the estimates. To understand the structures behind these differences and evaluate potential estimation biases with each index, I conducted factor analysis using both single-factor and multi-factor models. I found that the latent factor structure underlying the indices may be more complex than the standard congeneric model, indicating the existence of substantial measurement error correlation between indices and the existence of two factors instead of one. The final estimated common factor model was used to analyze the reliability, precision, and accuracy of each index. Notably, I found that the GPT Beta index from Eloundou et al. (2023) is consistently estimated to be the measurement with the highest reliability ratio with regard to the modeled latent factor. In addition, I constructed a composite exposure index that maximizes correlation with the latent factor by extracting the factor scores from the fitted common factor models, providing an alternative index that aggregates information from multiple exposure indices for future researchers to deploy. Using the extracted composite exposure index, I examine the distribution and characteristics of Occupational AI Exposure in the U.S. labor market. Consistent with previous studies, I found that occupations with higher annual mean wages and workers with higher education levels, on average, face higher AI exposure. I further applied the composite exposure index to the 3 empirical estimation settings and found that higher occupational AI exposure is associated with longer weekly work hours, larger occupational total employment, lower occupational annual mean wage, and higher worker unemployment risk.