Nearly 90% of Biomedical Papers Show Signs of AI-Assisted Writing, Study Finds
核心洞察
A preprint study estimates that almost nine out of ten papers published in December 2025 in a major biomedical database showed signs of AI-assisted writing.
The rate of LLM usage reached 77% for papers published in 2025 and 52% for 2024, suggesting AI use in scientific writing is rising rapidly.
Researchers attribute the higher figures to a more sensitive detection method that yields direct estimates rather than lower bounds.
The use of artificial intelligence to write scientific papers could be far more prevalent than previously thought, according to a preprint study estimating that almost nine out of ten papers published in December 2025 in a major biomedical-article database showed signs of AI-assisted writing.
The study, posted on the arXiv preprint site on 12 August and not yet peer reviewed, puts the rate of usage of AI large language models (LLMs) at 77% for papers archived in the PubMed Central repository and published in the whole of 2025, and 52% for those published in 2024 — suggesting LLM use is on the rise. The study included only papers written in English.
These figures are substantially higher than previous estimates of LLM use in the scientific literature. A 2025 paper authored by some of the same researchers that analysed abstracts of papers in PubMed, rather than the full text, put the figure at at least 13.5% for 2024. Meanwhile, a 2026 study that looked at papers across academic disciplines estimated that 57% of 2025 papers were probably AI-influenced.
A More Sensitive Detection Method
Dmitry Kobak (搜索), a computer scientist at Ghent University in Belgium and co-author of the preprint and the 2025 paper, said he was initially sceptical of the calculations in the latest study: "I was sure that we did something wrong." But further checks convinced him that the data stood up.
Like many previous studies, the latest work analysed the frequency of words commonly used by LLMs to detect signs of their deployment in papers. Kobak attributes the higher figure to the particular method his team used this time around, which is more sensitive to LLM use and therefore more likely to give a higher figure. This method, which yields direct estimates for AI use rather than lower bounds, increased the LLM use rate for 2024 abstracts from 13.5% to 31%.
The authors also say that the figures are consistent with the findings of a survey conducted in 2025 — in which 71% of researchers said they use AI for writing assistance — and that the true figure is probably higher than people admit or report in surveys.
Quantity Over Quality
A separate modelling study, posted on the arXiv preprint repository on 19 July and also not yet peer reviewed, predicts that scientists who use LLMs to help with their research will spend less time refining their work and instead jump quickly to fresh projects. Adoption of LLMs will cause scientists to "do more, less well — rather than the same amount, better," the authors write.
The results reflect a flawed incentive system in science that prioritizes quantity over quality, according to study co-author Carl Bergstrom (搜索), a biologist at the University of Washington in Seattle. "LLMs are rarely the problem themselves," Bergstrom said. "LLMs hold up a mirror to problems that we already have."
To predict how LLMs will change scientific productivity, the authors broke the research process into discrete phases: a discovery phase, followed by a two-part development phase consisting of required work (such as creating figures and drafting papers) and discretionary development (such as conducting follow-up experiments and polishing writing). Drawing on methods from optimal-foraging theory, the model assumed LLMs are functioning at their best — cheap, fast and accurate.
The modelling predicts that LLMs can speed up all phases of the scientific process, but that this acceleration won't result in better papers. Faster discovery and required-work phases mean that researchers can churn out papers more quickly than they could without the help of LLMs, and the pressure to publish means there is little incentive to spend extra time polishing those papers.
A Lasting Shift
Other researchers told Nature that the high rates of estimated LLM use reported in the latest paper could make sense given the widespread use of LLMs. They also cautioned that the figures in the study might not be representative of the entire scientific literature and that more analysis is needed to fully understand the rates of usage more broadly.
But the results indicate that LLMs are here to stay, said Kyle Siler (搜索), a social scientist at the University of Toronto, Canada, and author of the 2026 study that estimated lower LLM use. "The toothpaste is out of the tube, and it's not going back."