Matteo Caruso (University of Turin)
This paper evaluates the early labour-market impact of Generative Artificial Intelligence (AI) on the United States using the public release of ChatGPT in November 2022 as a natural experiment. I combine IPUMS ACS micro-data (20212024) with a novel, revealed-preference measure of occupational AI exposure built from the Anthropic Economic Index (AEI, Handa et al. 2025), which classifies millions of real Claude conversations into O*NET task categories and distinguishes between automation (full delegation) and augmentation (iterative use) interactions. Starting from this task-level evidence, I aggregate exposure at the occupation level and group the 974 Job-Explorer occupations into three mutually exclusive categories Automation-dominant, Augmentation-dominant, and Irrelevant used as the treatment structure in a di?erence-in-di?erences design. I find no significant e?ect of Chat-GPT exposure on aggregate employment, a positive wage e?ect of around three log-points for Automation-dominant occupations, and an increase in usual weekly hours of roughly 0.50.9 hours in both exposed groups. Women gain disproportionately from the augmentation channel across wages, employment, and hours. The results are consistent with an interpretation in which workers adopted ChatGPT privately to accelerate their most automatable tasks and translated the time saved into higher productivity rather than displacement.