Scientists have used writing-assistance software for years, but large language models have changed the scale of the question: it’s no longer just whether a paper was polished with software, but how much of its prose an AI model actually generated. When a team of researchers set out to measure that shift across the biomedical literature, the results surprised even them. “I was sure that we did something wrong,” Dmitry Kobak, a computer scientist at Ghent University, told Nature.
Kobak and his team detailed their findings in a preprint posted to arXiv, analyzing word-frequency patterns characteristic of large language models across biomedical papers archived in PubMed Central.1 The share of papers showing these AI-writing signatures—word frequency patterns that indicate AI-assisted writing—climbed from 52 percent in 2024 to 77 percent across all of 2025, reaching nearly 90 percent by December 2025.
What the Data Shows: AI-Writing Signatures Nearly Doubled in a Year
The share of biomedical papers in PubMed Central showing word-frequency signatures characteristic of large language models rose from 52 percent in 2024 to nearly 90 percent by December 2025, according to a preprint analysis by Holzwarth, González-Márquez, and Kobak.
Adapted from Holzwarth, et al, 2026. Image generated using Claude.
By the Numbers
|
The signatures were not evenly distributed across a paper's sections. Discussion sections were the most likely to show AI involvement, at 78 percent, compared with 58 percent of results sections in the same December 2025 papers. That distinction matters, Kobak said, because AI involvement in a results section carries different risks than in a discussion: large language models have a well-documented tendency to fabricate, or “hallucinate,” data and citations, a failure mode with much higher stakes than AI-polished prose in a paper's closing argument.
Why It Matters: Peer Review Wasn't Built for an AI-Saturated Literature
The trend also troubles researchers who study how science communicates with itself. “The toothpaste is out of the tube, and it's not going back,” Kyle Siler, a sociologist at the University of Toronto, told Nature. Most journals still rely on authors to voluntarily disclose AI use, an honor-system approach that has produced mixed results so far. With AI-writing signatures now the norm rather than the exception in biomedical publishing, editors and reviewers may need tools built for a literature where assuming a paper is “AI-free” is no longer a reasonable default.
Use of AI Declaration: This article was produced under The Scientist’s AI Editorial Policy.
- Holzwarth L, et al. Most biomedical publications show signs of LLM-assisted writing. arXiv. 2608.10715.
- Kobak D, et al. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Sci Adv. 2025;11(27):eadt3813.
