| Flagged abstracts | Pangram flagged 29.4 percent of U.S. STEM dissertation abstracts filed through May 2026 as AI-generated, up from 1.7 percent in 2023, according to NBER Working Paper 35859. |
|---|---|
| Industry gap | Among U.S.-origin graduates of the most research-intensive universities who finished from 2023 to 2025, a flagged abstract lined up with a 7.4 percentage-point lower chance of an academic job in 2026 and an 8.0 percentage-point higher chance of industry work. |
| Campus rates | In the 2026 filings, 41 percent of abstracts from students educated in non-English-speaking countries were flagged, against 23 percent from English-speaking countries. Computer science was 48 percent. |
| Not a cause | The authors call the job pattern a correlation, not proof that AI writing caused the career split. The paper has not been peer reviewed. |
Daniel P. Gross, Dror Shvadron and Hansen Zhang report that AI-written PhD dissertations made up 29.4 percent of U.S. STEM abstracts filed through May 2026. Their National Bureau of Economic Research paper, which has not been peer reviewed, also ties those dissertations to industry jobs rather than academic posts. Nature's report on the findings ran on Friday, Oct. 16, 2026.
AI-written PhD dissertations reach 29.4 percent
Pangram flagged 29.4 percent of U.S. STEM dissertation abstracts filed through May 2026 as AI-generated, the authors write in NBER Working Paper 35859. STEM means science, technology, engineering and mathematics. Nature's opening line said roughly 30 percent of dissertations "this year" contained such text. The paper counts abstracts, and the 2026 data stop in May.
Gross, an associate professor of business administration at Duke's Fuqua School of Business in Durham, N.C., wrote it with Dror Shvadron of the University of Toronto's Rotman School of Management and Hansen Zhang of Fuqua. NBER lists Gross as a research associate.
The abstracts come from ProQuest Dissertations and Theses Global and from university repositories, for graduates from January 2019 through May 2026. The appendix counts 226,911 graduates and 226,591 abstracts scored with Pangram 3.3.2. The authors say that tracks the National Science Foundation's Survey of Earned Doctorates, with nearly complete coverage of the life, physical and mathematical sciences and for about 85 percent of engineering.
Six of about 126,000 abstracts were flagged before 2023, under a 0.01 percent false-positive rate. The share then ran 1.7 percent in 2023, 5.5 percent in 2024 and 18.8 percent in 2025. In 2026, mathematical and computer sciences were at 36 percent, engineering at 35 percent, life sciences at 26 percent and physical sciences at 22 percent. Computer science was 48 percent.
From 340 full dissertations, the authors project AI writing in about 8 percent of 2024 texts and 31 percent in 2026, and they call the abstract rate a lower bound. Gross told Nature that 90 percent of flagged abstracts were classed as entirely machine-written. That figure is not in the paper.
Where the flags are highest
Students with earlier schooling in a non-English-speaking country were flagged more often in 2026: 41 percent, against 23 percent from English-speaking countries, the paper says. The flagged share was 44 percent for students from China, 43 percent from India, 49 percent from South Korea and 51 percent from Iran, against 23 percent for a U.S. bachelor's degree. LinkedIn supplies that country for about 60 percent of the sample.
Before 2023, the non-English group was flagged at the same zero rate as others, on about 18,000 graduates a year. The authors say that weighs against detector bias. Under the 2021 Carnegie Classification, 28 percent of 2026 dissertations at R1 universities were flagged, 31 percent at R2 schools and 37 percent at other doctoral schools. R1 is the highest research-activity group. Top-ten programs in a National Research Council ranking were at 15 percent in 2025 and 2026, and programs ranked 90 to 100 were at 26 percent. Among R1 schools with at least 50 graduates since 2025, MIT was at 9 percent, Caltech at 10 percent and Princeton at 13 percent.
Industry jobs and what the gap does not prove
U.S.-origin graduates with a flagged abstract were less often in academia in 2026 and more often in industry, the paper finds. The authors say that does not show AI writing caused the jobs.
They compare 2023 through 2025 graduates from the same R1 university and field, using LinkedIn histories from Revelio Labs. Of the 226,911 graduates, 70.2 percent matched a profile. For people with a U.S. degree before the doctorate, a flagged abstract meant 7.4 percentage points less academic work, on a 44 percent base, and 8.0 percentage points more industry work, on a 50 percent base. Nature stated those gaps as 17 percent and 16 percent in relative terms. Government employment did not differ, and the gaps shrank when the earlier degree was foreign.
Among graduates employed in academia, a flagged abstract meant a 5.4 percentage-point higher chance of a non-tenure-track post, on a 12 percent base. Tenure-track and postdoctoral posts showed no significant difference. Academic employers in the study include teaching hospitals and federally funded research centers.
Delegating the writing may leave students less prepared, or students already leaving research may write with less effort. Page counts do not track AI use, and the job gaps remain after that control. The authors write that they "thus far view the evidence as suggesting an adverse impact of AI use on research capability."
What a defense can and cannot test
The authors say a live defense may matter more if the chapter no longer shows who did the thinking. They also say the tools can help writers who are not working in a first language.
Gross told Nature, "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? Ang Liu, a professor of engineering design at the University of New South Wales, told Nature the hard case is research where knowledge is incomplete. "That is where rigorous research training becomes even more important," he said.
The acknowledgements say Claude Fable 5.1 drafted the discussion and part of the data appendix, and that the authors revised it by hand. Pangram gave API credits and a discount.
Comments
Post a Comment