> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI writing traces show up in nearly nine in ten biomedical papers from December 2025
- URL: https://www.metatalks.ai/ai-writing-traces-show-up-in-nearly-nine-in-ten-biomedical-papers-from-december-2025/
- Published: 2026-08-25T11:37:00.000Z
- Updated: 2026-08-25T11:36:59.000Z
- Author: Al
- Tags: News, Science AI, AI Detection, #newswire

**The work is a preprint that has not been peer reviewed — and the traces reach 58% of results sections, the part of a paper where invented data would do the most damage.**

Close to nine in ten papers published in December 2025 in PubMed Central, a large database of biomedical articles, carry traces of writing done with AI assistance, a study estimates.

Across full years, the same work puts large language models behind 77% of the repository's papers published across 2025, against 52% for 2024.

The study went up on the preprint server arXiv on 12 August and has not been through peer review. As in much of the earlier research, it looks for LLMs by counting how often words the models favour turn up in a text, and it examined papers written in English only.

Those shares run well above earlier estimates. [A 2025 paper](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list%5Fuids=40601754&ref=metatalks.ai) written in part by several of the same researchers looked at PubMed abstracts rather than whole texts and arrived at a floor of 13.5% for 2024\. [A separate 2026 study](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Abstract&list%5Fuids=42213755&ref=metatalks.ai), covering work from a range of scholarly fields, put the share of 2025 papers likely shaped by AI at 57%.

Dmitry Kobak, a computer scientist at Ghent University in Belgium who helped write both the preprint and the 2025 paper, told Nature his first reaction to the new numbers was doubt: "I was sure that we did something wrong." Further verification, he said, persuaded him the figures held up. He attributes the size of the figures to the technique his team used this time, which registers LLM use more readily and produces straight estimates instead of floor values; run over the same 2024 abstracts, it lifted the projected share from 13.5% to 31%.

Within a paper, the traces are not spread evenly. They show up more often in abstracts, introductions and discussions than in methods and results portions. Among December 2025 papers, an estimated 78% of discussion sections carried signs of AI use, compared with 58% of results sections. Kobak said results sections produced with AI help could be a cause for concern, since the models are prone to making up — hallucinating — data.

The authors say their numbers line up with a 2025 survey in which 71% of researchers reported turning to AI for help with their writing, and that the real rate is probably higher than what people admit or report. Other researchers told the journal that estimates that steep were plausible, with LLMs now so widely used. They also warned that the paper's figures may not stand for scientific publishing as a whole, and that pinning down how often LLMs are used more broadly will take further analysis.

Kyle Siler, a social scientist at the University of Toronto who wrote the 2026 study with the lower estimate, told the journal the findings indicate one thing all the same: LLMs are a permanent fixture, the toothpaste squeezed out of the tube and not going back in.