AI Is in 29% of New US STEM Dissertations. The Study That Found It Used AI Too.

Nearly three in ten US STEM PhD dissertations filed in 2026 contain AI-generated writing, according to a National Bureau of Economic Research working paper by Daniel P. Gross and Hansen Zhang of Duke's Fuqua School of Business and Dror Shvadron of the University of Toronto. The study ran the abstracts of about 226,000 STEM dissertations from 2019 to May 2026 through an AI detector and linked the graduates to their LinkedIn careers. AI writing was effectively absent before 2023 and reached 29.4% this year, and graduates who used it were less likely to end up in academic research. Then there is the line in the acknowledgements: the authors used Claude Fable 5.1 to draft part of the paper, too.

From Six Flagged Abstracts to Nearly One in Three

The team pulled dissertations from ProQuest and university repositories, a sample that lines up closely with the National Science Foundation's annual census of US doctorates. They scored every abstract with Pangram 3.3.2, a commercial AI detector, and checked a stratified sample of full dissertations to confirm that an AI-written abstract was a fair proxy for AI writing in the body.

The detector's false alarms are the first thing a skeptic would ask about, and the data answer it neatly. Across roughly 126,000 dissertations filed between 2019 and 2022, written before chatbots like ChatGPT were in wide use, Pangram flagged six abstracts in total. That works out to a false positive rate under 0.01%.

After that, the curve bends hard: 1.7% of 2023 dissertations, 5.5% in 2024, 18.8% in 2025 and 29.4% in the 2026 cohort. Because the test only looks at abstracts, the authors treat these as lower bounds and project that about 31% of 2026 dissertations contain AI writing somewhere. At the current growth rate, they write, a majority of US STEM dissertations will contain AI-generated content by 2028.

The study was released this month and covered by Nature on October 16. It has not been peer-reviewed.

Civil Engineering at 51%, MIT at 9%

Use is uneven. In the 2026 cohort, AI-generated abstracts showed up in 36% of mathematical and computer science dissertations, 35% in engineering, 26% in the life sciences and 22% in the physical sciences. By narrower field, the range runs from 17% in mathematics and statistics to 48% in computer science and 51% in civil engineering. Chemistry and physics sit near the bottom, which the authors link to abstracts full of notation, compounds and quantities.

Language is the biggest divider. Among 2026 graduates, 41% of dissertations by students educated in non-English-speaking countries had AI-written abstracts, against 23% for students from English-speaking countries. By origin, the rate was 51% for students from Iran, 49% from South Korea, 44% from China and 43% from India. Students from Canada and the UK looked no different from US-origin students.

Prestige matters too. At R1 universities, 28% of 2026 dissertations contained AI writing, compared with 31% at R2 schools and 37% at other doctoral institutions. Among R1 universities with at least 50 graduates since 2025, the rate was 9% at MIT, 10% at Caltech and 13% at Princeton, while the ten highest-rate universities were above 30%. Some top programs, including mathematics at Berkeley, had no flagged abstracts at all.

The Career Gap the Authors Can't Fully Explain

The most provocative result is what happened after graduation. Looking at US-origin graduates of R1 universities from 2023 to 2025, the authors compared people in the same university and field. Those with AI-written dissertations were 7.4 percentage points less likely to work in academia in 2026, a 17% drop from a 44% base rate, and 8.0 points more likely to work in industry. Among those who stayed in academia, they were 5.4 points more likely to hold non-tenure-track jobs. For likely foreign-origin graduates, the differences were much smaller.

The paper is explicit that this is correlation, not cause. One reading is that writing with AI leaves students with a training gap. The other is reverse causation: students who already plan to leave research may put less effort into the dissertation. The authors tested one sign of low effort, dissertation length, and found it didn't track AI use, and the career gaps held when they controlled for it. On that basis they say the evidence so far suggests "an adverse impact of AI use on research capability," while admitting it is too soon to know for sure.

That fits a pattern WikiDigit has covered before. A Berkeley-led experiment found that ten minutes of AI help made people quit sooner once the chatbot was taken away, especially when they asked for answers instead of hints. And the developer behind an AI-found 3D "einstein" tile has said he can follow only the intuition behind the proof his model produced.

Yes, the Paper Used Claude

The acknowledgements say Claude Fable 5.1 "was used in the initial drafting of the discussion of the empirical findings, and for a subset of the data appendix," and that all AI-generated text was substantially revised by hand. That's not hypocrisy so much as a demonstration of the paper's own point. Experienced economists can hand a first draft to a model and still judge what comes back. The worry the study raises is about students who never build that judgment because the model wrote the hard parts before they learned how.

The authors also give the strongest counterargument themselves. Writing science in English is costly for non-native speakers, and earlier research found that generative AI helps weaker writers the most. AI could narrow gaps in writing quality while also removing the careful thinking that writing forces. Their data fit both readings.

Their suggested safeguard is old technology: the live dissertation defense. Unlike journals, PhD programs already make candidates explain their work in person, and the paper argues that test may matter more now. As the authors put it, whether the next generation of scientists is "trained or merely credentialed remains to be seen."

Sources

• NBER Working Paper 35859: How Heavily Are PhD Students Offloading Their Scientific Writing to AI? (Gross, Shvadron, Zhang)
• NBER Working Paper 35859 (PDF)
• Nature: PhD students who write dissertation with AI less likely to stay in academia
• The Chronicle of Higher Education: Would You Hire a Ph.D. Who Wrote Their Dissertation With AI?