Census linked graduates: AI-exposed majors show worse first jobs, with sector downgrading

Cody Orr, Lee C. Tucker and Lawrence Warren, “Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors,” U.S. Census Center for Economic Studies working paper CES-26-56 (September 2026); full original. Career-path brief. Administrative university and earnings/industry records for graduates in covered states provide unusually useful evidence on actual first employment, wages and initial job destinations rather than just posting counts. Authors compare exposed majors across cohorts surrounding the November 2022 release of ChatGPT, adjusting for covariates and pre-period differences, not assigning individual employers to adopt AI at random.

Most-exposed decile of majors versus their less-exposed counterparts: regression-adjusted likelihood of initial observed employment about five percentage points lower; full-quarter initial earnings among employed graduates about 13% lower. Approximately half the pay decline is lower earnings within employing sectors and half is movement toward lower-paying sectors such as restaurants/retail. Initial declines diminish with time since graduation but remain material for the most exposed cohorts at observed horizons. This adds an actual job quality mechanism missed by studies only counting headcount; it does not measure lifetime wage loss or AI-caused layoffs.

Limit: exposure of a college major is a predicted occupational-task measure, not a record that the hiring employer used AI or rejected a specific applicant because of it. Closely timed differences in industry demand, remote-work rules, tech hiring or composition of graduates could produce similar patterns; author event-study/pretrend checks reduce but cannot eliminate this. Unemployment outside covered linked records, later wage recovery and young people who never enrolled in college remain open. Conditioning earnings on observed full-quarter employment changes the composition compared with all graduates. Pair with Fairlie and Wu’s population unemployment check and remote-work alternative, not as if the three have the same exposure definition.