AI-written dissertations hit 29% of US STEM PhDs, NBER paper finds

Just over 29% of US STEM PhD dissertations submitted in the first five months of 2026 contained AI-generated writing, up from 1.7% in 2023 and effectively zero before that. The figure comes from a new National Bureau of Economic Research working paper covered by Nature, which examined abstracts from more than 226,000 STEM doctoral graduates at US universities between January 2019 and May 2026.

The authors, Daniel P. Gross of Duke University with Dror Shvadron and Hansen Zhang, used the detection service Pangram to flag abstracts; roughly 90% of flagged abstracts were classed as entirely AI-generated.

Then they followed where the graduates went. Among those who had done their undergraduate degrees in the US, graduates whose abstracts showed AI assistance were 17% less likely to be employed in academia and 16% more likely to work in industry than peers who did not use AI. Employment was tracked through LinkedIn profiles.

Usage is uneven across the dataset. Mathematical and computer sciences lead at 36%. Dissertations by graduates from non-English-speaking countries contained AI writing 41% of the time, against 23% for those from English-speaking countries. The authors also report heavier use at lower-ranked PhD programs.

Gross frames the stakes in old-fashioned terms. He told Nature that "putting things into words forces a degree of precision" and that "the writing is thinking." The paper's abstract puts it more clinically: "Large language models are diffusing rapidly into science, raising concerns that delegating research writing to AI may undermine the development of expertise that scientific training is supposed to build."

Ang Liu, at the University of New South Wales, said AI tends to help where substantial existing knowledge already exists, but gets in the way of the "rigorous research training" demanded in uncertain research areas. The paper does not resolve whether AI use is pushing graduates away from academia or whether industry-bound students were simply more willing to use it in the first place.