Nearly 30% of U.S. STEM PhD Abstracts Included AI-Generated Text in 2026

Almost 30% of abstracts from U.S. science, technology, engineering and mathematics (STEM) doctoral dissertations completed in the first five months of 2026 contained text generated by artificial intelligence, according to a new research paper.¹

The study reviewed dissertation abstracts submitted to online databases from 2019 through May 2026. It found that doctoral graduates whose papers included AI-generated text were more likely to work in industry and less likely to enter academia.

The findings come from a working paper by researchers at the National Bureau of Economic Research (NBER), a U.S.-based nonprofit organization. The paper has not been peer-reviewed.

AI use in PhD abstracts rose sharply after 2023

Several studies have reported a significant increase in AI-generated text in scientific literature in recent years. The NBER researchers wanted to examine how doctoral students are using AI because student training and performance could shape the future of scientific research.

“If students are only doing half the work and letting AI do the rest, do they really have the deep understanding to be good scientists?” asks co-author Daniel Gross, an economist at the NBER in Durham.

The researchers analyzed more than 226,000 STEM doctoral dissertations completed at U.S. universities between January 2019 and May 2026. They used the AI-detection service Pangram to assess the abstracts and categorized them according to whether they contained AI-generated text.

The analysis did not distinguish between abstracts written entirely by AI and those written partly with AI assistance. However, Gross says that Pangram classified 90% of the abstracts flagged as AI-generated as completely AI-generated.

AI-generated text first appeared in the dataset’s dissertation abstracts in 2023, when it was identified in 1.7% of abstracts. That share rose to 18.8% in 2025 and 29.4% among abstracts submitted during the first five months of 2026.

AI use was highest in mathematics and computer science, where 36% of abstracts included AI-generated text. About 41% of abstracts by graduates educated in non-English-speaking countries included AI-generated text, compared with 23% of abstracts by graduates from English-speaking countries.

Researchers question the impact on scientific training

Anne Liu, an engineer at the University of New South Wales in Sydney, Australia, says generative AI can create problems when doctoral students are working in areas where the answers are not yet known.

AI can be useful when students are working in fields with a substantial existing body of knowledge, Liu says. But when research enters uncertain or incomplete areas, rigorous research training becomes even more important.

“That’s where rigorous research training becomes even more important,” she says.

Doctoral graduates using AI were more likely to enter industry

The researchers also examined doctoral graduates’ academic and employment profiles on LinkedIn. Among U.S. doctoral graduates with a bachelor’s degree, those whose abstracts included AI-generated text were 17% less likely to work in academia than graduates whose abstracts did not contain AI-generated text.

At the same time, graduates whose abstracts included AI-generated text were 16% more likely to work in industry. The relationship between AI use and academic employment was less pronounced among graduates who completed their undergraduate studies outside the United States.