An August 11 preprint analyzing 1.19 million open-access biomedical papers in PubMed Central estimated that 89% of papers published in December 2025 showed an excess of vocabulary associated with large-language-model assistance. The authors describe an estimate of assisted-writing prevalence, not proof that AI generated an entire paper, with implications for publication-policy debates.
A new preprint estimates that large-language-model assistance may have left a measurable imprint on most open-access biomedical papers published near the end of 2025. The study, posted on arXiv on August 11 by Lena Holzwarth, Rita González-Márquez and Dmitry Kobak, analyzed 1,194,287 full-text papers in PubMed Central.
What the preprint estimates
The authors track changes in the frequency of words that became more common after the widespread availability of large language models. Their method compares observed vocabulary with a counterfactual trend from earlier years, then estimates the share of papers showing excess LLM-associated vocabulary.
For papers published in December 2025, the estimate was 89%. The paper reports that the signal was more common in Discussion paragraphs than in Methods paragraphs: 68% versus 32% in length-matched samples. The authors also report that more than half of Methods sections showed the signal when assessed as a whole. That figure should not be read as a count of papers written entirely by AI. It is a statistical estimate of LLM-associated vocabulary in the study's open-access PubMed Central corpus. It cannot, on its own, establish how much assistance any individual author used or whether that assistance affected the underlying research.
Why the distinction matters
Nature's reporting on the preprint similarly describes the result as evidence of AI-assisted writing, not a direct authorship determination. The paper says its estimates could help inform guidelines and policies as journals and research institutions decide how to document AI use.
For people building tools on scientific literature, the result underscores a practical provenance question: writing assistance may be widespread even when the scientific claims, data and analysis still require their own evaluation. The preprint remains unreviewed research, so its methodology and interpretation should be assessed on that basis.
Key Points #
- 1The August 11 arXiv preprint analyzed 1,194,287 open-access biomedical papers from PubMed Central.
- 2It estimated that 89% of papers published in December 2025 showed excess LLM-associated vocabulary.
- 3The estimate is a corpus-level signal of assisted writing, not proof that AI generated an entire paper or that an individual paper used AI.
Scoring Rationale #
A large-scale preprint offers a timely, method-specific estimate of LLM-assisted writing in biomedical publishing, with direct implications for research provenance and editorial policy.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.