Recent findings indicate that a substantial number of biomedical papers exhibit characteristics of AI-generated text. Image credit: ©iStock.com, Nobi_Prizue While researchers have long utilized software to refine their writing, large language models have fundamentally altered the scope of the issue. The question is no longer merely whether a paper has been subtly edited, but rather the extent to which an AI model has produced the actual prose. When a group of researchers sought to quantify this shift across the biomedical literature, the findings were even unexpected to them. “I was sure that we did something wrong,” stated Dmitry Kobak, a computer scientist at Ghent University, in an interview with Nature.
Kobak and his colleagues published their findings in a preprint on arXiv, examining word-frequency patterns indicative of large language models within biomedical papers archived in PubMed Central. The proportion of papers exhibiting these AI-writing signatures—specific word-frequency patterns that suggest AI-assisted writing—increased from 52 percent in 2024 to 77 percent throughout all of 2025, approaching nearly 90 percent by December of that year.
What the Data Shows: AI-Writing Signatures Have 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.
Key Statistics
- 52 percent of biomedical papers displayed AI-writing signatures in 2024, a figure that rose to 77 percent across 2025.
- Nearly 90 percent of papers published in December 2025 exhibited AI-writing signatures.
- Discussion sections were the most impacted (78 percent), compared to 58 percent of results sections.
- An initial abstract-only estimate placed 2024 figures at 13.5 percent, which increased to 31 percent when the team reapplied its refined methodology to abstracts alone.
The signatures were not uniformly distributed across the various sections of a paper. Discussion sections demonstrated the highest likelihood of AI involvement, at 78 percent, compared with 58 percent of results sections in the same December 2025 papers. Kobak emphasized that this distinction is critical because AI involvement in a results section presents different risks than its presence in a discussion. Large language models have a well-documented tendency to fabricate, or “hallucinate,” data and citations—a failure mode carrying far greater stakes than AI-polished prose in a paper’s concluding argument.
Why It Matters: Peer Review Was Not Designed for an AI-Saturated Literature
This trend also concerns researchers who study the communication of science. “The toothpaste is out of the tube, and it’s not going back,” remarked Kyle Siler, a sociologist at the University of Toronto, in an interview with Nature. Most journals currently depend on authors to voluntarily disclose AI usage, an honor-system approach that has yielded mixed results thus far. As AI-writing signatures become the norm rather than the exception in biomedical publishing, editors and reviewers may require tools designed for a literature where assuming a paper is “AI-free” is no longer a reasonable default.

