Roughly nine in ten biomedical papers now show signs of AI-assisted writing, a preprint estimates

The finding

A preprint posted to arXiv on 11 August estimates that roughly 89 percent of biomedical papers published in December 2025, and archived in the open-access PubMed Central repository, carry vocabulary signatures of large language model assistance. Across the whole of 2025 the figure is 77 percent, up from 52 percent for papers published in 2024. Nature reported the study on 20 August under the headline that nine in ten biomedical papers now show signs of AI help.

The work is a preprint by Holzwarth, González-Márquez and Kobak, Most biomedical publications show signs of LLM-assisted writing. It has not been peer reviewed. The authors analysed the full text of 1,194,287 English-language papers published between 2017 and 2025, tracking 379 function and style words whose usage rose sharply across the literature after ChatGPT was released, and reading the statistical excess of those words as a marker of machine-assisted writing or editing.

What the number is, and what it is not

The method detects a vocabulary fingerprint, not authorship. A high score means a paper reads as though a language model helped write or edit it. It is not evidence that AI wrote the paper, that anything was fabricated, or that any rule was broken. Much of what it captures is ordinary editing of prose by researchers who use these tools the way earlier cohorts used spellcheckers and grammar tools.

The estimate is also higher than earlier ones because the method is more sensitive, not only because usage rose. An earlier paper from some of the same authors put LLM signs in 2024 biomedical abstracts at at least 13.5 percent; the more sensitive approach in the new work raises the 2024 abstract figure to about 31 percent. A separate 2026 study by Kyle Siler at the University of Toronto estimated 57 percent of 2025 papers across disciplines were probably AI-influenced. The figures are proxies that disagree by wide margins, which is itself worth holding onto.

The gradient that matters

The signal is not evenly spread through a paper. The study finds LLM markers roughly twice as common in discussion paragraphs, about 68 percent, as in methods paragraphs, about 32 percent. Dmitry Kobak of Ghent University, a co-author, notes that assistance concentrated in introductions and discussions is where a model's own tendencies can quietly shape how a field frames its questions, and that assisted results sections would be the more troubling case given the known tendency of these tools to fabricate.

Analysis

For a platform that weighs evidence rather than counts it, the headline percentage is the least useful part of this. Two things follow that are. First, "shows signs of AI writing" and "is unreliable" are different claims, and conflating them would discredit a large and mostly ordinary body of work while distracting from the narrow case that actually warrants scrutiny, machine-generated results and citations. Second, the section gradient is a better guide to where to look than any global rate: a polished discussion tells you little, while unusual phrasing in methods or results is worth a second read. The proxies also disagree enough that no single figure should be quoted as settled.

What to watch

Whether the preprint survives peer review with its estimates intact, whether detection methods that rely on word frequency keep working as models are tuned away from their tell-tale vocabulary, and whether journals move from disclosure policies toward section-aware checks that treat an AI-assisted discussion differently from AI-generated data.