A preprint study posted to arXiv on 12 August 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 analysis covered 1.19 million English-language papers and found LLM usage rates of 77% for all of 2025 and 52% for 2024, up sharply from 19% in 2023.
The work, led by Dmitry Kobak of Ghent University and colleagues, uses a word-frequency method that yields direct estimates rather than lower bounds. That methodological shift raised the 2024 abstract-level estimate from 13.5% in an earlier study to 31% in the new analysis. The authors say the figures align with a 2025 survey in which 71% of researchers reported using AI for writing help, and that actual use is likely higher than surveys capture.
What's new
- December 2025 snapshot: ~90% of PubMed Central papers showed AI writing signals.
- Full-year 2025: 77% of papers; 2024: 52%; 2023: 19%.
- Section variation: Discussion sections led at 78% (December 2025), followed by abstracts (68%), introductions (63%), results (58%), and methods (54%).
- Regional gap: Non-English-speaking countries such as South Korea and China exceeded 80% AI use, versus lower rates in the UK and US, likely driven by language-polishing needs.
- Method: Frequency analysis of LLM-favored words; produces direct estimates, not lower bounds.
- Status: Preprint (arXiv:2608.10715), not yet peer reviewed.
Why it matters
The concentration of AI signals in discussion sections raises the risk that model biases — rather than evidence — shape the narrative framing of entire fields. Results sections, where hallucinated data could enter the record, still showed a 58% signal rate. Meanwhile, a companion modelling study (arXiv:2507.14234) argues LLMs accelerate the discovery and drafting phases but reduce time spent on discretionary polishing, predicting a shift toward "more, less well" output under current publish-or-perish incentives.
Our take
The 90% figure is arresting, but it reflects a more sensitive detector, not necessarily a sudden surge in misconduct. The real story is the section-level pattern: discussions and abstracts — the parts readers skim most — carry the strongest AI imprint. If journals and funders want disclosure policies to mean anything, they will need section-specific guidelines and verification tooling, not blanket bans.
