AI-Assisted Writing Across Medical Journals: Heterogeneity in Prevalence and Trends
- 1Brown University
Abstract
BackgroundEstimates of generative artificial intelligence (AI) use in the medical literature vary widely, potentially reflecting differences across journals. We estimated changes in the prevalence of AI-generated writing over time using abstracts from 12 general medical journals.
MethodsWe randomly sampled up to 200 abstracts per journal-year from publications dated January 1, 2020, through August 8, 2026. Abstracts were scored using GPTZero, with journal-specific thresholds calibrated to fix a 1% false-positive rate (FPR) in the pre-AI baseline period (2020-2021). We estimated annual detector-positive proportions and calculated a publication-volume-weighted estimate across sampled journals. Associations with journal impact factor, publication model, and publication volume were examined descriptively.
ResultsAmong 15,268 abstracts, estimated prevalence of AI-generated text increased after 2022, with substantial variation in timing and magnitude across journals. In 2026 through August 8, prevalence estimates ranged from 5.7% (95% CI, 2.1%-10.0%) in the New England Journal of Medicine to 83.5% (95% CI, 77.1%-87.5%) in Frontiers in Medicine. The publication-volume-weighted estimate across sampled journals was 66.7% (95% CI, 64.0%-69.2%), and trends were insensitive to specification of false-positive rate in baseline data. The proportion of abstracts showing signs of AI-generated text was lower in journals with higher impact factors and higher in fully open-access journals.
ConclusionsEstimated prevalence of AI-generated text increased substantially across all sampled medical journals, with marked between-journal differences associated with journal impact factor and publication model. These findings support development of standardized, transparent AI reporting in medical publishing.
Related
- Author Gender Inequality in Medical Imaging Journals and the COVID-19 Pandemic
- From Discovery to Verification: A Thirty-Three-Year Analysis of Research Design in Three Leading General Medical Journals, 1992–2024
- University-managed publishing of medical journals: Reflections on the transfer of three journals from a commercial publisher to Kashan university of medical sciences
- Attitudes and perceptions toward the use of artificial intelligence chatbots for peer review in medical journals: a large-scale, international cross-sectional survey.
- Feasibility of Privacy-preserving Locally Hosted Language Models for Assessing Medical Student Reflective Writing.
BibTeX
@misc{flanders2026aiassisted,
title = {{AI-Assisted Writing Across Medical Journals: Heterogeneity in Prevalence and Trends}},
author = {M. D. Flanders and M. Caunca and A. Mehrotra and S. F. Ackley},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.24.26366055},
url = {https://www.medrxiv.org/content/10.64898/2026.10.24.26366055v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - AI-Assisted Writing Across Medical Journals: Heterogeneity in Prevalence and Trends
AU - M. D. Flanders
AU - M. Caunca
AU - A. Mehrotra
AU - S. F. Ackley
PY - 2026
DA - 2026/10/24
PB - medRxiv
DO - 10.64898/2026.10.24.26366055
UR - https://www.medrxiv.org/content/10.64898/2026.10.24.26366055v1
AB - BackgroundEstimates of generative artificial intelligence (AI) use in the medical literature vary widely, potentially reflecting differences across journals. We estimated changes in the prevalence of AI-generated writing over time using abstracts from 12 general medical journals. MethodsWe randomly sampled up to 200 abstracts per journal-year from publications dated January 1, 2020, through August 8, 2026. Abstracts were scored using GPTZero, with journal-specific thresholds calibrated to fix a 1% false-positive rate (FPR) in the pre-AI baseline period (2020-2021). We estimated annual detector-positive proportions and calculated a publication-volume-weighted estimate across sampled journals. Associations with journal impact factor, publication model, and publication volume were examined descriptively. ResultsAmong 15,268 abstracts, estimated prevalence of AI-generated text increased after 2022, with substantial variation in timing and magnitude across journals. In 2026 through August 8, prevalence estimates ranged from 5.7% (95% CI, 2.1%-10.0%) in the New England Journal of Medicine to 83.5% (95% CI, 77.1%-87.5%) in Frontiers in Medicine. The publication-volume-weighted estimate across sampled journals was 66.7% (95% CI, 64.0%-69.2%), and trends were insensitive to specification of false-positive rate in baseline data. The proportion of abstracts showing signs of AI-generated text was lower in journals with higher impact factors and higher in fully open-access journals. ConclusionsEstimated prevalence of AI-generated text increased substantially across all sampled medical journals, with marked between-journal differences associated with journal impact factor and publication model. These findings support development of standardized, transparent AI reporting in medical publishing.
ER -