An analysis of the PubMed Central corpus detects markers of LLM use in 77% of articles in 2025, compared to 52% a year earlier — well above previous estimates.
A preprint submitted on August 12, 2026 and relayed by Nature measures the presence of stylistic markers characteristic of major language models in articles deposited on PubMed Central. The result far exceeds previous estimates: 77% of articles published in 2025 bear these traces, compared to 52% for those of 2024.
The distribution within the articles is instructive. Markers are much more frequent in abstracts, introductions and discussions than in the methods and results sections — in other words, in the editorial sections, where assisted writing saves time without touching the data itself.
The method has its known limitations: it identifies regularities of style, not an admission of usage, and some of the texts detected are probably a proofreading or a translation aid rather than a complete generation. Nevertheless, it gives an order of magnitude that is difficult to ignore for journals that now require a declaration of AI use, and which, for the most part, have no means of verifying it.
Read the full article on the original source
Go to the source →