NBER: 29% of 2026 STEM PhD abstracts flagged for AI

Twenty-nine percent of US STEM PhD dissertation abstracts filed through May 2026 carried AI-generated text, up from essentially zero before 2023, according to a new NBER working paper covered by Nature.

The paper, 'How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations,' is authored by Daniel P. Gross of Duke's Fuqua School of Business with Dror Shvadron and Hansen Zhang. The authors ran roughly 226,000 US STEM dissertations completed between January 2019 and May 2026 through the AI-detection service Pangram, then linked each graduate to their first job. The year-over-year climb: 1.7% in 2023, 5.5% the next year, 18.8%, and 29.4% in the first five months of 2026.

Use is not evenly distributed. Math and computer science abstracts flagged at about 36%. Dissertations by graduates from non-English-speaking countries flagged 41% of the time, versus 23% for English-speaking peers. Students at lower-ranked programs showed higher rates too.

The career finding is what makes this more than a plagiarism story. Among US-origin graduates whose abstracts the detector flagged, PhDs were 17% less likely to pursue academic employment and 16% more likely to enter industry than same-cohort, same-program peers without flagged AI use. The pattern was strongest among US-origin students.

'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?' Gross asks.

The caveats are blunt. The paper is a working paper, not peer-reviewed. Pangram's flags are a probabilistic detector output, not a confession, and the authors score the abstract rather than the full dissertation. Whether students using AI are self-selecting toward industry or being pushed out of academic tracks by weaker research skills is not something a correlational design can separate.