NBER Study Finds AI Writing in 29% of US STEM PhD Dissertations, and Those Grads Are Less Likely to Stay in Academia

Nearly three in ten US STEM PhD dissertations filed this year contain AI-generated writing, according to a new NBER working paper by Daniel P. Gross and Hansen Zhang of Duke University and Dror Shvadron of the University of Toronto. The exact figure is 29.4% for dissertations filed through May 2026. In 2023 it was 1.7%.

The paper also links graduates to their LinkedIn job histories. US-origin graduates whose dissertations contain AI writing are less likely to be working in academia and more likely to be in industry. The authors say clearly that this is a correlation, not proof that AI hurt anyone’s training.

The study is trending on X, and The Chronicle of Higher Education covered it on October 5 (US time). “There’s no sign of this slowing down,” Gross told the Chronicle.

The numbers, year by year

The researchers ran the abstracts of roughly 226,000 US STEM dissertations filed between 2019 and May 2026 through Pangram 3.3.2, a commercial AI text detector. The sample comes from ProQuest plus university repositories and closely matches the NSF’s Survey of Earned Doctorates, they say.

Graduation year	Share of dissertations with AI-generated abstracts
2019 to 2022	Effectively 0% (6 flags out of about 126,000)
2023	1.7%
2024	5.5%
2025	18.8%
2026 (through May)	29.4%

Those pre-2023 numbers double as a test. Students who wrote before ChatGPT could not have used it, so six flags out of 126,000 suggests a false positive rate below 0.01%, the authors argue.

They also treat the figures as a floor, not a ceiling. Adjusting for AI use that shows up in the full text but not the abstract, they project about 31% for 2026. At the current pace, they write, a majority of US STEM dissertations will contain AI-generated content by 2028.

Where AI writing shows up most

Adoption is uneven by field. Among 2026 graduates, AI-generated abstracts appear in 36% of math and computer science dissertations, 35% in engineering, 26% in the life sciences and 22% in the physical sciences.

The spread inside those groups is wider. Computer science sits at 48% and civil engineering at 51%. Mathematics and statistics is at 17% and chemical engineering at 18%. The authors note that low-use fields such as chemistry and physics tend to have abstracts packed with notation, compounds and quantities.

Computer science sitting near the top fits a wider pattern. Researchers already use models for open math problems, and some software teams have gone much further, like 37signals going “pencils down” on hand-written code.

Who is using it

The biggest split is language. Among 2026 graduates, 41% of dissertations by students educated in non-English-speaking countries had AI-generated abstracts, compared with 23% for students from English-speaking countries. The paper counts the US, Canada, the UK, Australia, New Zealand and Ireland as English-speaking.

By country of origin, the paper reports 51% for students from Iran, 49% from South Korea, 44% from China and 43% from India, against 23% for students with US lower degrees. Students from Canada and the UK did not differ from US-origin students.

Prestige matters too. In 2026, 28% of dissertations at R1 research universities had AI writing, compared with 31% at R2 schools and 37% at other doctoral institutions. Among 2025 and 2026 graduates of top-ten programs in their field, the rate was 15%.

Among R1 universities with at least 50 graduates since 2025, the AI share was 9% at MIT, 10% at Caltech and 13% at Princeton. Mathematics at Berkeley and chemistry at Ohio State had no AI-generated abstracts at all.

The career link, and why it is only a correlation

This is the part drawing the most debate online. The researchers looked at where 2023 to 2025 graduates of R1 universities were working in 2026 and compared students in the same university and field.

Among US-origin graduates, those with AI-written dissertations were 7.4 percentage points less likely to be employed in academia, against a 44% base rate. They were 8.0 points more likely to be in industry, against a 50% base. There was no difference in government jobs.

Within academia, graduates with AI-written dissertations were 5.4 points more likely to hold non-tenure-track roles, on a base of 12%. For likely foreign-origin graduates, the differences shrank substantially.

The authors offer two readings. One is that writing with AI leaves a training gap that makes a research career harder. The other is reverse causality: students who have already decided to leave academia may simply put less effort into the dissertation.

They lean toward the first reading, carefully. Dissertation length does not track AI use, and the job gaps hold after controlling for it. “We thus far view the evidence as suggesting an adverse impact of AI use on research capability,” they write. They also say it is “likely too soon to know for sure.”

How much should you trust the detector?

AI detectors have a bad reputation, and older ones were known to flag non-native English writing unfairly. The authors address that directly. Before 2023, abstracts by students from non-English-speaking countries were flagged at 0%, the same as their English-speaking peers, on samples of roughly 18,000 graduates per year. The gap only appears after ChatGPT arrived.

There are still limits. The main measure looks at abstracts, not full dissertations. The full-text check covers 340 randomly sampled dissertations, which is a small base for validation. And the paper discloses that Pangram supplied API credits and discounted pricing.

One more detail stands out. The authors disclose that Claude Fable 5.1 helped draft the paper’s discussion of findings and part of its data appendix, with all AI text “substantially manually revised.” Even a paper worried about AI writing used it, and said so.

What the paper shows	What it does not show
29.4% of 2026 US STEM dissertation abstracts are flagged as AI-generated	That 29.4% of dissertations were mostly written by AI
AI use is higher among non-native English speakers and at lower-ranked programs	That those students are less capable researchers
US-origin graduates with AI writing are less often in academia and more often in industry	That AI use caused those career outcomes
A near-zero false positive rate on pre-2023 abstracts	How well the detector handles heavily edited AI text
A projection that most dissertations will contain AI writing by 2028	Any change to university rules or degree requirements

What it means for Indian developers

India is one of the biggest sending countries for US STEM PhD programs, so the 43% figure for Indian-origin graduates is personal for a lot of families. Comparing students inside the same programs, the paper puts the extra AI use by Indian-origin students at about 11 percentage points in 2025 and 2026.

The paper does not call that a bad thing on its own. The authors cite research showing AI helps writers who face the biggest language costs, and say AI may “close gaps in writing quality or comprehension while also removing the impetus for the careful thinking which writing requires.” Both can be true at once.

For anyone heading into a US PhD, or hiring PhD graduates in Bengaluru, Hyderabad or Pune, the practical point is simple. A polished abstract proves less than it used to. The live defense, which the authors note may grow in importance, and the ability to explain your own work without notes are what will separate people.

The career data is murkier for Indian students than the headlines suggest. The academia versus industry gap was clearest for US-origin graduates and much smaller for foreign-origin ones. And industry is hardly punishing AI skills: companies are spending heavily to train engineers who work alongside AI.

Our take

The headline number is solid. A near-census of dissertations and a clean pre-ChatGPT baseline are hard to argue with. The career finding is where to slow down: it is a LinkedIn correlation, and reverse causality is plausible.

Still, the question the paper raises is a fair one. “Putting things into words forces, for me, a degree of precision that otherwise might have gone missing,” Gross told the Chronicle. “The writing is thinking, as people like to say.” If that is right, the cost will not show up in this year’s dissertations. It will show up in what this cohort can do five years from now.

FAQ

Who wrote the study on AI in PhD dissertations?

Daniel P. Gross and Hansen Zhang of Duke University’s Fuqua School of Business and Dror Shvadron of the University of Toronto’s Rotman School of Management. It is NBER Working Paper 35859, issued in October 2026. Like other NBER working papers, it has not been through peer review.

Does 29% mean AI wrote 29% of dissertations?

No. It means 29.4% of 2026 dissertations had abstracts that the Pangram detector flagged as AI-generated. The paper treats this as a signal of AI writing in the dissertation, not a measure of how much of each thesis was machine-written.

Which fields use AI writing the most?

Among 2026 graduates, civil engineering (51%) and computer science (48%) were at the top of the range. Mathematics and statistics (17%) and chemical engineering (18%) were among the lowest.

Are AI detectors biased against non-native English speakers?

Older detectors were found to be. In this study, abstracts written before ChatGPT by students from non-English-speaking countries were flagged at 0%, the same as native speakers, which the authors use to argue that Pangram does not show that bias here.

Does using AI in a dissertation hurt your career?

The study cannot say that. US-origin graduates with AI writing were less likely to work in academia and more likely to work in industry, but the authors stress the link is correlational and could reflect students who had already planned to leave research.

Featured image: PhD theses at ETH Zurich. Photo by Easyloc via Wikimedia Commons, licensed CC BY-SA 4.0. Cropped and resized.