NBER: AI-written PhDs 17% less likely to stay in academia

Nearly 30% of US STEM PhD dissertation abstracts filed in the first five months of 2026 contained AI-generated text, and the graduates behind them were 17% less likely to pursue academic jobs than peers whose abstracts showed no AI, according to a National Bureau of Economic Research working paper covered by Nature.

The paper, by economist Daniel Gross with co-authors Dror Shvadron and Hansen Zhang, ran the Pangram AI-detection service across more than 226,000 doctoral abstracts filed between January 2019 and May 2026. AI-flagged text first surfaced in 2023 at 1.7% of abstracts, climbed to 18.8% in 2025, and reached 29.4% in the 2026 cohort. Math and computer science led the fields at 36%. Dissertations by graduates from non-English-speaking countries were flagged 41% of the time, versus 23% for English-speaking ones.

For US-origin PhDs with AI-flagged abstracts, the tilt ran away from universities. They were 17% less likely to take academic jobs and 16% more likely to go into industry, with the correlation weaker for those educated outside the US.

Gross, based at NBER in Durham, told Nature he questioned whether students relying heavily on AI "really have the depth of understanding" needed for scientific careers. Ang Liu, an engineer at the University of New South Wales, said that when researchers push into areas with incomplete knowledge, "rigorous research training becomes even more important," pointing at the limits of what current AI can do at the frontier.

The paper measures text, not intent: a flagged abstract could reflect anything from heavy ghostwriting to light copy-editing, and the authors publish no breakdown by use case. Two researchers we track shared the Nature piece this week.