NBER: PhDs who wrote with AI less likely to stay in academia

The share of US STEM PhD dissertations with AI-generated text in the abstract has climbed from 1.7% in 2023 to 29.4% in the first five months of 2026. That is from NBER Working Paper 35859, by economist Daniel P. Gross with co-authors Dror Shvadron and Hansen Zhang, who ran more than 226,000 dissertations through the AI-detection service Pangram. Roughly 90% of flagged abstracts were classified as entirely AI-generated. Nature's writeup by Mohana Basu lays out the field split: mathematical and computer sciences at 36%, physical sciences at 22%.

The career finding sits in the US-origin subsample. PhDs there whose abstracts flagged as AI-written were 17% less likely to take academic jobs and 16% more likely to go to industry.

Gross is blunt about what that implies: "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?

Rates vary sharply by background too. Among graduates from non-English-speaking countries, 41% of abstracts flagged, versus 23% for graduates from English-speaking countries. Ang Liu, an engineer at the University of New South Wales in Sydney, told Nature that AI works best with established knowledge and struggles in frontier research areas: "That is where rigorous research training becomes even more important."

The paper measures abstracts, not full dissertations. Pangram is a classifier, not a confession, and the 17-point academic-employment gap is a correlation drawn from one working paper.