Signs of artificial intelligence use appear more often in a paper’s introduction and discussion than in its results section.Credit: Lawrence Dutton/Getty
Artificial intelligence may be used to help write scientific papers far more often than previously estimated. A new study suggests that nearly nine in ten articles published in December 2025 and indexed in a major biomedical database showed signs of AI-assisted writing.
The study was published on the arXiv preprint site on August 121. Although the research has not yet been peer reviewed, it estimates that large language models (LLMs) were used in 77% of papers archived in the PubMed Central repository and published during 2025. That figure compares with 52% of papers published in 2024, suggesting that AI use in scientific writing is increasing. The analysis included only papers written in English.
These estimates are considerably higher than those reported in earlier studies of AI use in scientific literature. A 2025 paper by some of the same researchers analyzed PubMed abstracts rather than full texts and estimated that at least 13.5% of papers published in 2024 showed evidence of LLM use.2 Meanwhile, a 2026 study covering multiple academic disciplines estimated that 57% of papers published in 2025 were likely influenced by AI.3
Dmitry Kovac, a computer scientist at Belgium’s Ghent University and a co-author of both studies, said he was initially skeptical of the latest analysis. He was “convinced we had done something wrong”, but further testing led him to conclude that the results were reliable.
As in many previous studies, the researchers looked for words and phrases commonly associated with LLM-generated text. Kovac attributes the higher estimates to the more sensitive detection method used in the latest analysis. The approach is intended to provide a direct estimate of AI use rather than a lower bound. Using this method, the estimated share of abstracts showing signs of LLM assistance in 2024 rose from 13.5% to 31%.
The authors say their findings are also consistent with a survey conducted in 2025. The survey found that 71% of researchers use AI to assist with writing. The actual proportion could be higher, because researchers may not always disclose or accurately report their use of AI tools.
Other researchers told Nature that the high levels of LLM use reported in the study could be plausible given the growing adoption of generative AI. However, they cautioned that the findings may not represent the entire scientific literature and that additional research is needed to establish broader usage rates. Kyle Siler, a social scientist at the University of Toronto in Canada and an author of the 2026 study estimating changes in LLM use, said the findings reinforce the idea that AI will remain part of scientific publishing.3 “The toothpaste comes out of the tube and doesn’t come back in.”
AI use varies by paper section
The study also found that LLM use was more common in abstracts, introductions and discussion sections than in methods and results sections. In papers published in December 2025, an estimated 78% of discussion sections showed signs of AI assistance, compared with 58% of results sections.
Kovac said the presence of AI-assisted writing in results sections could be concerning because LLMs can fabricate information or “hallucinate” data. AI-generated text in introductions may also shape how researchers describe the central ideas and established knowledge of a field. “Whatever prejudices LLMs have, they’re suddenly going to permeate the literature,” Kovac said.
Source: www.nature.com