NBER: STEM PhDs who wrote with AI leave academia more often

A working paper from the US National Bureau of Economic Research scanned more than 226,000 STEM dissertations filed at US universities between January 2019 and May 2026 and found AI-written text in 1.7% of abstracts in 2023, 18.8% in 2025, and 29.4% of abstracts filed in the first five months of 2026, according to Nature.

The authors ran the abstracts through the AI-detection service Pangram. Mathematical and computer sciences topped the fields, with about 36% of abstracts flagging. Graduates educated in non-English-speaking countries had AI-written abstracts 41% of the time, versus 23% for those from English-speaking countries.

The career finding is the one that will travel. Among US-undergraduate PhDs, those with AI-flagged abstracts were 17% less likely to go into academia and 16% more likely to go into industry than peers whose writing did not flag.

Daniel Gross, an economist at NBER, put the stake plainly: "If our students are only doing half the work and they're dishing the rest to AI, do they really have the depth of understanding" needed for rigorous science. Ang Liu, an engineer at the University of New South Wales, told Nature the harder problem comes "when research moves into areas in which knowledge is incomplete or uncertain."

The paper leaves the direction of causation open: the Nature write-up reports a correlation between AI-flagged writing and the industry route, not a demonstration that one caused the other.