Predictive AI Diffusion and Population Health: Evidence from U.S. Household Panel Data, 2011–2023

• In Event: Health IT and Outcomes

Abstract

Earlier waves of automation, from industrial robotics to trade-related displacement, carried measurable costs to worker health, and comparable harms have been widely anticipated for artificial intelligence (AI). We test this expectation for predictive AI---the classification, forecasting, and decision-support systems that diffused across United States workplaces through the 2010s. We use annual health reports from approximately 150,000 households in the NielsenIQ Consumer Panel Ailments module, 2011--2023, linked to metropolitan measures of AI-skill penetration derived from Revelio Labs résumé and LinkedIn data. Using household and year fixed effects together with a shift-share instrument built on year-2000 industrial composition, we estimate effects on nearly forty diagnosed physical and mental health conditions across 242 metropolitan areas. Across outcomes including insomnia, depression and anxiety, hypertension, and back pain, we find no evidence of statistically or economically significant effects. Point estimates are small and centered on zero, and the null holds across subgroups defined by age, income, education, and occupation, and under alternative measures of occupational AI exposure. Predictive AI thus diffused across the United States without the population-level health costs documented for earlier automation waves.

Author

• Youzheng Jia, University of Pittsburgh
  Presenting Author