Linking degrees to paychecks
Three economists at the U.S. Census Bureau, Cody Orr, Lee C. Tucker and Lawrence Warren, linked the Bureau's records of college graduates' early-career outcomes to employer earnings records. Their sample covers institutions awarding roughly 29 percent of U.S. bachelor's degrees from 2016 to 2024, about 6.7 million graduate records. They rank majors by how exposed the jobs their graduates usually take are to large language models, using task-based exposure measures from earlier research by Eloundou and colleagues, and compare the most exposed tenth of majors with the bottom six deciles before and after ChatGPT's release in late 2022. [1]
A recession-sized hit, mostly in computer science
Outcomes for the most exposed majors began to diverge immediately after ChatGPT arrived. In the authors' regression-adjusted estimates, graduates of the most exposed tenth became five percentage points less likely to be employed shortly after graduating, and their full-quarter starting earnings fell 13 percent, a loss comparable to graduating into a large recession. That decile is made up largely of computer science and related majors. [1]
About half of the earnings decline came from lower pay within the industries that hire these graduates. The rest came from more of them taking jobs in lower-paying sectors such as retail and food service. The gap narrowed as graduates moved further from entry, to about five percent after two years, but stayed substantial for the most exposed majors. [1][2]
What the paper does not settle
The authors say their short time series means they cannot yet assess when or whether exposed graduates catch up. The design assumes the most exposed majors and the comparison group would otherwise have moved together, and they note that pandemic-era shocks and the rise of remote work are not entirely controlled for, although results held when they controlled for each major's work-from-home rate. It is a working paper that has not had the review given to official Census Bureau publications. [1]
Atlas interpretation: The divergence lines up with ChatGPT, but it also lines up with the uneven recovery from the pandemic, which the authors say their design does not fully separate out. The paper is strong evidence that something changed for these graduates after late 2022, and suggestive rather than conclusive about how much of it AI caused. [1]