AI-written PhD abstracts linked to industry exit from academia

Nearly 30% of US STEM dissertation abstracts filed in the first five months of 2026 contained AI-generated text, up from 1.7% in 2023, according to a National Bureau of Economic Research working paper reported by Nature. The PhDs whose abstracts were flagged are also measurably less likely to end up in academia.

The working paper, by economist Daniel Gross with co-authors Dror Shvadron and Hansen Zhang, scanned more than 226,000 STEM dissertations completed between January 2019 and May 2026 using the Pangram AI-detection service. Prevalence climbed from 1.7% in 2023 to 18.8% in 2025, then 29.4% by May 2026. Mathematical and computer sciences led at 36%, engineering followed at 35%, and physical sciences trailed at 22%.

The split by language background is wider. Among graduates from non-English-speaking countries, 41% of abstracts carried AI-generated text, versus 23% for graduates from English-speaking countries.

Among US-educated PhDs, those with AI-flagged abstracts were 17% less likely to pursue academic careers and 16% more likely to work in industry than peers without the flag. The effect was less pronounced for internationally-educated graduates.

Gross put the worry plainly to Nature: "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 go into a scientific career? Ang Liu of the University of New South Wales told the magazine that when research explores uncertain areas, "rigorous research training becomes even more important."

Two things sit unresolved in the setup. Pangram's error rate on non-native-English scientific prose is not reported here, and the 41 / 23 split suggests the detector may be catching some ESL writing patterns alongside actual AI use. The paper also does not establish direction of causality: students already heading for industry careers may simply invest less in prose craftsmanship.