NBER study ties AI-written dissertations to leaving academia

Roughly 30% of PhD dissertations in science, technology, engineering and mathematics completed at US universities this year contained AI-generated text, Nature reports, citing a working paper that scanned the abstracts of more than 226,000 STEM doctoral graduates between January 2019 and May 2026. Among those who completed their undergraduate degrees in the United States, graduates with AI-generated abstracts were 17% less likely to be employed in academia than those who did not use AI, and 16% more likely to be employed in industry.

The jump is recent. AI-detected text went from 1.7% of abstracts in 2023 to 18.8% in 2025 to 29.4% through the first five months of 2026. Mathematical and computer sciences lead the fields at 36%. About 41% of dissertations by graduates educated in non-English-speaking countries were flagged, compared with 23% from English-speaking countries.

The study is a National Bureau of Economic Research working paper by Daniel Gross and colleagues, and has not been peer reviewed. The authors used the AI-detection service Pangram to classify abstracts as AI-written or not, without distinguishing wholly from partly generated text. Gross told Nature that 90% of the flagged abstracts were classified by Pangram as entirely AI-generated.

"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" to be good scientists, asked Gross, an economist at NBER. Ang Liu, an engineer at the University of New South Wales in Sydney, pointed to novel research as the pressure point: "That is where rigorous research training becomes even more important."