The estimate rose sharply through 2025
The preprint analyzes word-frequency shifts in more than one million English-language open-access papers in PubMed Central. It estimates that 52 percent of 2024 papers and 77 percent of 2025 papers showed signs of LLM assistance, reaching 89 percent among papers published in December 2025.
Nature reports that discussion sections in December were estimated at 78 percent and results sections at 58 percent. The authors argue that their method is more sensitive than earlier work and produces direct prevalence estimates rather than a conservative lower bound.
The number does not identify an AI author
The method detects vocabulary patterns associated with model use. It cannot reconstruct what task a model performed in each paper, whether the model edited grammar, drafted a section, summarized literature, analyzed data, or introduced an error.
The study is a preprint, covers a specific open-access biomedical corpus, and has not yet undergone peer review. Other researchers told Nature that the estimates could be plausible given widespread tool use while warning that the corpus may not represent all scientific publishing.
Disclosure should follow contribution, not style
Using a language tool to improve clarity is different from delegating interpretation of results, generating citations, or writing claims the authors did not verify. A binary detector cannot govern those distinctions reliably.
Journals and institutions should require task-specific disclosure, preserve human responsibility for every claim and citation, audit higher-risk uses in methods and results, and make correction pathways fast. The goal is not to punish polished prose. It is to keep the scientific record traceable and trustworthy as assistance becomes routine.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
Nature — Biomedical papers show widespread signs of AI help
