90% of Biomedical Papers Now Show Signs of AI-Assisted Writing

Posted on 24 August, 2026 by Mehrdad Fathi

A preprint posted to arXiv on 12 August 2026 estimates that nearly nine out of ten biomedical papers published in December 2025 and archived in PubMed Central show signs of AI-assisted writing. The figure is far above earlier published estimates, and it reframes a question the research community has been treating as speculative: large language model (LLM) use in scientific writing is no longer an emerging trend to monitor. It is the default condition of the literature.

The trajectory in the numbers

The study, which has not yet been peer reviewed and covers English-language papers only, reports a clear year-over-year climb:

• 52% of papers published in 2024
• 77% of papers published across the whole of 2025
• Roughly 90% of papers published in December 2025

For context on how sharply this diverges from prior work: a 2025 paper by some of the same researchers, analysing PubMed abstracts rather than full texts, put the 2024 figure at at least 13.5%. A separate 2026 study spanning multiple academic disciplines estimated that 57% of 2025 papers were probably AI-influenced.

Dmitry Kobak, a computer scientist at Ghent University in Belgium and a co-author of both the new preprint and the 2025 paper, was initially sceptical of his own team’s result: “I was sure that we did something wrong.” Further checks held up.

Why the number jumped

The methodology explains most of the gap. Like earlier work, the team tracked the frequency of words that LLMs use disproportionately. The difference is that this approach produces direct estimates of AI use rather than lower bounds, making it more sensitive to lighter-touch assistance. Applied retroactively to 2024 abstracts, the new method lifted the estimate from 13.5% to 31%.

The authors argue the findings are consistent with self-reported behaviour: a 2025 survey found 71% of researchers using AI for writing assistance, and they suspect the real rate exceeds what respondents disclose.

Where in the paper AI shows up matters

The section-level breakdown is the most practically useful part of the analysis. AI signatures cluster in abstracts, introductions and discussions rather than methods and results. For December 2025 papers, an estimated 78% of discussion sections showed signs of AI, compared with 58% of results sections.

Both figures carry distinct risks. AI-assisted results sections are the more acute concern given LLMs’ tendency to fabricate data. AI-assisted introductions carry a subtler, compounding one: if models shape how researchers frame the state of a field, model bias propagates through the literature. As Kobak puts it, “Whatever bias the LLM may have will just suddenly permeate the literature.”

The detection problem nobody has solved

Current detection methods cannot distinguish between text an LLM generated and text an LLM merely corrected for grammar, notes Andrew Gray, a bibliometrics support officer at University College London and author of a 2025 study that estimated lower 2024 usage. That limitation has direct operational consequences for publishers: a paper written wholesale by a model and a paper polished for readability produce similar linguistic fingerprints, so screening submissions for low-quality AI content on lexical signals alone is unreliable.

The equity dimension

LLM use was more common among authors based in countries where English is not the predominant language. Kobak considers this expected — non-native speakers routinely use these tools to translate and refine text. It complicates any blanket policy response, since the same tool that enables low-effort output also lowers a long-standing barrier to publication. Kyle Siler, a social scientist at the University of Toronto and author of the 2026 cross-disciplinary study, argues that the more valuable research question now is not who uses LLMs, but how.

What this means going forward

Many journals now require authors to disclose how and when AI was used. Disclosure addresses provenance, not quality. Lena Holzwarth, the study’s first author and a cognitive scientist who worked on it at the Eberhard Karls University of Tübingen, makes the case that the drafting process itself has scientific value: wrestling with ideas and committing them to language is part of how thinking gets tested, and outsourcing it removes that check.

Siler’s assessment of where this lands is blunt: “The toothpaste is out of the tube, and it’s not going back.” The open question for research organiz longer whether to permit AI-assisted writing, but which uses strengthen a manuscripses strengthen a manuscript and which quietly erode its reliability.

Other researchers cautioned that these figures, drawn from a biomedical repository, may not represent the whole of the scientific literature, and that broader analysis is needed.

Source: Adapted from Nature news, doi: https://doi.org/10.1038/d41586-026-02551-z

References:

1. Holzwarth, L., González-Márquez, R. & Kobak, D. Preprint at arXiv https://doi.org/10.48550/arXiv.2608.10715 (2026).

2. Kobak, D., González-Márquez, R., Horvát, E.-Á. & Lause, J. Sci. Adv. 11, eadt3813 (2025).

3. Siler, K. Proc. Natl Acad. Sci. USA 123, e2605754123 (2026).

4. Gray, A. Preprint at arXiv https://doi.org/10.48550/arXiv.2512.01560 (2025).