Overview
- After authors first adopted chatbots, preprint output rose by more than 36% on arXiv, roughly 53% on bioRxiv, and nearly 60% on SSRN.
- AI-assisted abstracts became more linguistically complex, yet those manuscripts were less likely to be accepted by peer-reviewed journals, reversing the usual complexity–quality signal.
- Productivity gains were most pronounced for researchers facing language barriers, with authors bearing Asian names at institutions in Asia nearly doubling submissions on bioRxiv and SSRN and rising by over 40% on arXiv.
- LLM adopters cited a broader mix of sources, including more books, younger works, and less-cited documents, indicating a diversification of referenced literature.
- The authors caution that AI use may be underdetected and that publication lags could bias acceptance measures, as related evidence of rising objective errors at NeurIPS heightens concern about peer-review capacity.