NBER working paper puts AI-generated text at 29.4% of early-2026 US STEM dissertation abstracts

The unit measured is the abstract, not the dissertation. According to Nature's report, the researchers took abstracts submitted to an online database ^[1], ran them through the Pangram detector, and labelled each one AI-written or not, with no separate category for partly generated text ^[4]. Co-author Daniel Gross said 90% of the flagged abstracts were classed by Pangram as entirely AI-generated ^[5]. That is the detector's call, and the full dissertations were not scored.

The trend is steep. Flagged abstracts rose 17.1 percentage points between 2023 and 2025 ^[17] and another 10.6 points to the 2026 figure ^[18]. But the 2026 number covers only dissertations filed in the first five months of the year ^[7], so setting it beside 2025's full-year 18.8% ^[6] assumes that January to May looks like the rest of the year.

Rates differ by field and by background. In mathematical and computer sciences, 36% of abstracts were flagged ^[8]. The share was 41% for graduates educated in non-English-speaking countries and 23% for those from English-speaking countries, a gap of 18 points [c9, d3].

The job result covers a subset: graduates who did their undergraduate degrees in the US. Among them, flagged abstracts went with a 17% lower chance of academic employment and a 16% higher chance of industry employment ^[10]. For graduates who studied as undergraduates elsewhere, the link to an academic job was less pronounced ^[11]. The team read employment from LinkedIn profiles ^[15]. Field is a candidate confounder, given the 36% rate in mathematical and computer sciences ^[8].

The thing this doesn't tell you is whether the gap holds within a field, what baseline employment rates the 17% and 16% apply to, or how often Pangram labels a human-written abstract as AI.

In our view the trend is the firmer finding. It depends on Pangram's labels being roughly right. The job comparison depends on that and on a control for who chooses to use AI.

The authors frame the question around training. The NBER team set out to study how PhD students use AI on the view that what students learn in training, and what they can do as a result, will shape science for decades to come ^[16]. Gross, an economist at NBER based in Durham, said: "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 ^[12].

Ang Liu, an engineer at the University of New South Wales in Sydney, said AI can be extremely helpful where a substantial body of knowledge already exists, and that the challenge comes when research moves into areas where knowledge is incomplete or uncertain ^[13]. "That is where rigorous research training becomes even more important," he said ^[14].