Artificial intelligence transparency statements and publication governance in medical literature: cross-sectional analysis of 97,916 articles, 2015–2026

Abstract

Large language models (LLMs) are increasingly used across scholarly writing, peer review, and editorial workflows, creating new research integrity questions about attribution, accountability, and disclosure. We measured how explicit artificial intelligence (AI) transparency statements are appearing in published medical literature and whether scalable full-text methods can monitor their adoption across journals and specialties. We conducted a retrospective cross-sectional bibliometric analysis of 97,916 English-language medical journal articles published from 2015 through April 2026 across 14 medical and surgical fields. Full text was retrieved from PubMed Central XML or publisher-hosted article files and classified with a large language model for positive AI-use disclosure, explicit statements of non-use, named tools, use cases, and statement location. Validation used pre-ChatGPT specificity, second-model agreement in 10,979 articles, and blinded human review of 300 stratified cases by 2 independent reviewers with adjudication. All rates were recalculated excluding Frontiers journals, whose submission workflow applies a standardized AI declaration. Any AI transparency statement increased from 1.4% in 2023 to 22.4% in January-April 2026, or 9.9% excluding Frontiers journals. Positive disclosure reached 7.3% (6.2% excluding Frontiers), while negative disclosure reached 15.1% (3.7% excluding Frontiers). Frontiers contributed 6.1% of articles but 71.1% of negative disclosures. Health Informatics and Digital Medicine had the highest positive disclosure rate during 2023–2026 (11.1%; 10.5% excluding Frontiers). Among 91 journals with at least 20 classified articles in 2025–2026, statement rates ranged from 0 to 100%, and all 3 journals at 100% were Frontiers titles. Language refinement was the most common disclosed use (72.8%). Inter-model agreement was high (Cohen kappa = 0.96 positive, 0.98 negative), and blinded human review confirmed 93.3% and 97.3% of these classifications, respectively. AI transparency statements are becoming a measurable publication-governance signal, but their prevalence is driven substantially by publisher submission infrastructure rather than author initiative: one publisher accounted for 71.1% of all negative disclosures. Near-universal declaration capture is therefore achievable through workflow design, and low observed disclosure rates may reflect missing declaration fields more than author reticence. Scalable full-text monitoring may help editors, publishers, and research integrity stakeholders evaluate policy implementation and distinguish assistive from generative AI use.