AI-generated dissertation text is now common in STEM, and the career path appears to differ
Roughly 30% of STEM PhD dissertations filed in the United States this year contained AI-generated text, a rise that would have been difficult to imagine only a few years ago. The pattern is not just about writing speed or convenience. In this dataset, graduates whose dissertation abstracts were flagged as AI-written were more likely to enter industry and less likely to move into academia after graduation.
The findings come from a working paper that reviewed abstracts submitted to an online dissertation database from 2019 through May 2026. It is not peer reviewed, but the scale is hard to ignore: more than 226,000 STEM doctoral graduates were included. Researchers used the AI-detection service Pangram to classify dissertation abstracts as either AI-written or not, without separating fully generated text from partially generated text. One striking detail emerged from the detection process itself: about 90% of the flagged abstracts were classified by Pangram as entirely AI-generated.
The rise was abrupt. AI-generated text first appeared in dissertation abstracts in 2023, then climbed from 1.7% that year to 18.8% in 2025. For dissertations filed in the first five months of 2026, 29.4% of abstracts contained AI-generated text. In mathematical and computer sciences, the share was even higher at 36%. Among graduates educated in non-English-speaking countries, about 41% of dissertations had AI-generated abstracts, compared with 23% among those from English-speaking countries.
A change that fast does not look like a passing anomaly. It looks like a structural shift in graduate training, publication habits and the expectations placed on early-career researchers.
What the dissertation data say about research training
The question at the center of the analysis is not whether generative AI can help polish prose. It clearly can. The deeper concern is whether heavy reliance on machine-written text changes what doctoral training is supposed to produce: independent reasoning, technical judgment and the ability to work through uncertain questions without a ready-made answer.
Daniel Gross, an economist at the National Bureau of Economic Research in Durham, framed the concern bluntly: if students are only doing part of the work and handing the rest to AI, it becomes harder to know whether they have developed the depth of understanding expected of future scientists. Ang Liu, an engineer at the University of New South Wales in Sydney, drew a related distinction. AI can be helpful when students are operating in fields with a substantial existing knowledge base. The pressure rises when research moves into uncertain territory, where the answer is not already in the training data and rigorous research training matters most.
That point reaches beyond style or authorship disputes. Dissertation writing is often where graduate students learn how to synthesize evidence, defend a conceptual framework and explain why an unanswered problem deserves attention. If AI is doing a meaningful share of that synthesis, the training signal may weaken even when the final document reads more smoothly.
Job pathways split between academia and industry
The employment analysis adds a second layer. Researchers also reviewed education and job profiles on LinkedIn. Among PhD graduates who completed their undergraduate degrees in the United States, those with AI-generated abstracts were 17% less likely to be employed in academia than graduates who did not use AI. At the same time, they were 16% more likely to be employed in industry.
For graduates who completed their undergraduate studies outside the United States, the academic association was less pronounced. The article does not explain why the pattern differed by educational background, and no causal claim is justified from these data alone. Even so, the direction of travel is difficult to miss. AI use in dissertation writing was associated with a career outcome that leaned away from academia and toward industry, at least in the subgroup of US-educated undergraduates.
The result raises a practical question for universities. If dissertation writing increasingly involves AI assistance, do graduate programs need to assess not just the final text but the process behind it? A polished abstract no longer guarantees that the student independently mastered the thinking that produced it.
Detection limits, classification choices and what remains uncertain
The methods deserve caution. Pangram was used to classify abstracts as AI-written or not, but the analysis did not distinguish between text that was partly generated by AI and text that was wholly generated by AI. That means the headline percentages are best read as an estimate of AI presence, not a precise measurement of how much of each dissertation was machine-produced.
The study also focused on abstracts, which are only a small part of a dissertation. Abstracts are the most visible and the most easily screened portion, but they may not reflect how AI was used elsewhere in the manuscript, or whether students relied on it for sentence-level editing, translation or full drafting. Still, abstract-level changes are meaningful because they are the first public-facing signal of how scholars present their work to the field.
There is also a clear institutional dimension. AI-generated abstracts were more common among graduates educated in non-English-speaking countries, a finding that may reflect language support needs as much as willingness to outsource thought. The difference does not prove a deficit in scholarship. It may instead reflect where students saw AI as a tool for expression rather than substitution.
The broader message is narrower and more concrete: dissertation writing is changing quickly, and doctoral training systems are already absorbing the consequences. Universities now face a test of their own. They can either treat AI-assisted writing as a marginal editing aid, or they can confront the possibility that graduate assessment, mentorship and definitions of authorship all need to be recalibrated for a generation that writes with machines in the loop.
The working paper does not answer how much AI use is acceptable, how universities should verify originality, or whether the career differences reflect hidden factors that were not measured. What it does show is that AI is no longer rare in STEM dissertation abstracts, and the graduates who use it are not moving through academia and industry in the same way as those who do not.