Staggering 90% of biomedical papers now show signs of AI help

A new study estimates that nearly nine out of ten biomedical papers published in December 2025 contained indications of being assisted by artificial intelligence. The research, posted on the arXiv preprint site on August 12 and not yet peer-reviewed, suggests that the use of large language models (LLMs) is on the rise.

According to the study, 77% of PubMed Central archived papers published in 2025 showed signs of AI assistance, with this figure increasing to 52% for papers published in 2024. These numbers are significantly higher than previous estimates, with a 2025 paper putting the rate at a minimum of 13.5% for 2024 and a 2026 study estimating 57% of 2025 papers potentially influenced by AI across various disciplines.

Dmitry Kobak, a computer scientist at Ghent University in Belgium and co-author of the preprint, initially found the results of the latest study to be implausible, but subsequent checks confirmed their validity. The researchers employed a more sensitive method to detect LLM use, which led to the higher estimate. They attribute the increased usage rate for 2024 abstracts from 13.5% to 31%.

The study also found that AI assistance was more prevalent in abstracts, introductions, and discussion sections compared to methods and results sections. Specifically, 78% of December 2025 discussion sections showed signs of AI, compared to 58% of results sections. Kobak warns that AI-assisted results sections could be problematic due to LLMs' tendency to fabricate data, known as 'hallucination'.

Additionally, AI-assisted writing of introductions could potentially skew the leading ideas within a scientific field by imparting any biases the models may have.

The authors of the study suggest that these findings are consistent with a 2025 survey, in which 71% of researchers reported using AI for writing assistance. The researchers believe that the true figure for LLM use in scientific literature might be even higher than what people admit or report in surveys. However, they caution that the study's figures might not be representative of the entire scientific literature and that further analysis is needed to fully understand the rates of usage more broadly.