AI-written dissertations are spreading fast across US STEM PhDs
Roughly 30% of STEM doctoral dissertations completed in the United States in early 2026 contained AI-generated text, a striking rise from near absence just three years earlier. The pattern matters because dissertation writing is supposed to reflect a student’s ability to frame uncertainty, defend methods, and explain results in original academic prose. When that work is increasingly delegated to generative tools, the concern is no longer only authorship. It extends to training quality, intellectual ownership, and the pipeline into research careers.
The analysis came from a working paper that examined dissertation abstracts submitted to an online database from January 2019 through May 2026. More than 226,000 STEM doctoral graduates were included, making this one of the largest looks yet at how generative AI is entering the scholarly record at the graduate level. Investigators used the AI-detection service Pangram to classify abstracts as AI-written or not; they did not separate fully generated text from text that was only partly generated with AI assistance. Even with that conservative framing, the growth curve was steep. AI-generated abstracts first appeared in 2023, rose to 1.7% that year, reached 18.8% in 2025, and climbed to 29.4% in dissertations filed in the first five months of 2026.
The shift was not uniform across fields. Mathematical and computer sciences showed the highest use, with AI-generated text appearing in 36% of abstracts. The uptake was also more common among graduates educated in non-English-speaking countries, where 41% of dissertations had AI-generated abstracts compared with 23% among those from English-speaking countries. Those differences may reflect the same pressure points that affect academic writing more broadly: language fluency, disciplinary conventions, and the temptation to use machine-written prose to accelerate a process that is already heavily time constrained.
AI is not merely changing how students draft sentences. It is changing how universities assess whether those students have internalized the scientific method.
Academic job placement moved in the opposite direction
The most consequential finding sits beyond text detection. Among PhD graduates who completed their undergraduate degrees in the United States, those with AI-generated dissertation abstracts were 17% less likely to be employed in academia after graduation than peers whose abstracts did not show AI use. At the same time, they were 16% more likely to be employed in industry.
Those associations do not prove causation, and the working paper does not claim that AI use itself forces graduates out of academia. But the pattern raises an immediate question about how doctoral training is being evaluated. If dissertation writing increasingly becomes a task that can be compressed or outsourced, then committees, advisers, and employers may be seeing a weaker proxy for the habits academia prizes most: persistence with ambiguous data, comfort with open-ended reasoning, and the ability to build an argument from partial evidence.
For graduates educated outside the United States, the relationship between AI use and academic employment was less pronounced. That difference may reflect variations in language support, institutional expectations, or labor-market pathways. It may also reflect how employers interpret writing quality when the baseline for academic English differs across training environments. The paper does not resolve those possibilities, but it does show that the labor-market signal around AI-assisted dissertation writing is not the same for every group.
The detection problem sits at the center of the debate
The use of Pangram as an AI-detection tool gives the findings their force, but also their vulnerability. Dissertations were categorized simply as AI-written or not, without a separate estimate of how much of each abstract came from a machine. Gross said 90% of the flagged abstracts were classified by Pangram as entirely AI-generated, yet the study still cannot tell whether students used AI for polishing, translation, drafting, or full composition. Those distinctions matter for interpretation, even if the paper treats the output in binary terms.
Still, the scale of the trend is hard to dismiss. Once AI-generated text appears in nearly one-third of dissertation abstracts in a national STEM sample, the issue is no longer a niche concern about a few opportunistic students. It becomes a structural feature of doctoral education. Universities are likely to face pressure to decide whether disclosure is enough, whether AI use should be restricted in dissertation writing, or whether training should shift toward methods that make students demonstrate comprehension in settings less vulnerable to automated text generation.
There is also a deeper scientific concern. Dissertation work is supposed to be where a trainee learns to live inside uncertainty. A model can produce fluent prose in seconds, but it cannot replace the intellectual strain of deciding what matters when the answer is not already known. In fields where evidence is incomplete or methods are still evolving, that strain is the training itself.
What the labor market may be signaling about AI-heavy training
The employment findings may reflect how employers read the broader package of skills attached to a dissertation, not just the document itself. Academic hiring rewards independence, originality, and the capacity to sustain a research agenda over years. Industry, by contrast, may value speed, applied problem-solving, and collaboration in environments where polished communication can be augmented by digital tools. If AI use is concentrated among students who are already optimizing for efficiency, the workplace pattern could reflect self-selection as much as skill erosion.
Even so, the practical consequence is straightforward. Graduate programs that allow AI to seep into dissertation writing without clear guardrails may be training students for a different labor market than the one universities think they are serving. The paper’s data suggest that AI adoption is already broad enough to influence outcomes after graduation, not just classroom writing habits.
The working paper remains unreviewed, and its conclusions are limited by the strengths and weaknesses of text-detection software. The classification system cannot capture every form of assistance, nor can it determine why a graduate moved into industry or stayed out of academia. But the central signal is difficult to ignore: AI use in dissertation writing has accelerated rapidly, and the graduates most likely to lean on it are not ending up in the same jobs as their peers. The unresolved issue is not whether AI will enter doctoral writing; it already has. The harder question is how much of dissertation training academia is willing to let it replace.