A landmark analysis published in Nature on August 20 has sent shockwaves through the scientific community: an estimated 90% of biomedical papers published in December 2025 and archived in PubMed show signs of AI-assisted writing, a figure dramatically higher than previous estimates.
The study, led by computer scientist Dmitry Kobak of Ghent University and colleagues, was posted as an arXiv preprint on August 12. It estimates that 77% of papers archived in PubMed Central for all of 2025 showed AI influence, rising from 52% in 2024. Previous estimates — which relied on abstract analysis alone — had placed the figure at 13.5% for 2024.
The dramatic jump stems partly from the team's improved detection methodology. While earlier studies analyzed only abstracts, this research examined full-text papers. The team's more sensitive method, which provides direct estimates rather than lower bounds, also contributed to the higher figure.
Critically, the study found that AI use was unevenly distributed across paper sections. An estimated 78% of discussion sections showed signs of AI, compared with 58% of results sections. This disparity raises particular alarm: AI-assisted results sections risk introducing fabricated or hallucinated data, while AI-written introductions could subtly skew the framing and leading ideas within entire fields.
Whatever bias the LLM may have will just suddenly permeate the literature, Kobak warned. The findings are consistent with a 2025 survey in which 71% of researchers admitted to using AI for writing assistance — and experts believe the true figure is likely higher than what researchers self-report.
The results suggest that large language models have become deeply embedded in academic publishing at a pace that has outstripped institutional oversight. The toothpaste is out of the tube, and it's not going back, said Kyle Siler, a social scientist at the University of Toronto who authored a separate 2026 study estimating lower but still substantial AI use in papers across academic disciplines.
Other researchers cautioned that the PubMed-based figures might not be representative of the entire scientific literature and called for further analysis. But the core message is clear: AI writing tools have become the norm, not the exception, in biomedical publishing — and the scientific establishment is only now beginning to reckon with the implications.
