Trends in artificial intelligence-generated and -assisted content in comment pieces in The Lancet Regional Health, 2020–2025: a longitudinal study and citation impact analysis

Abstract Background Large language models (LLMs) are now widely used in numerous text-generation exercises. This trend raises important questions about the scientific literature and the effectiveness of editorial checks. This study aimed to measure the frequency of artificial intelligence (AI)-written and assisted text in scientific comments over time and by regional journals. Methods We analysed the complete corpus (993 comment articles) published in The Lancet Regional Health (TLRH) journals between August 2020 and December 2025. We evaluated these texts using Pangram, an AI-detection tool. To test the detector itself, we used a random sample of 50 comments written on or before 2022, when LLMs became broadly available. We compared these original comments with our own versions assisted by ChatGPT and/or fully written by ChatGPT using the same topics. We also examined vocabulary patterns and grammatical details in four text groups. Finally, we compared the number of times comments classified by Pangram as human-written and AI-generated were cited in other papers and policy reports. Results The detector classified 100% of the 376 comments published on or before 2022 as human-written. Comments classified by Pangram as containing AI-generated text first appeared in 2023, representing 3.7% (7 of 185) of the total. This proportion rose to 11.3% (23 of 203) in 2024 and to 20.5% (47 of 229) in 2025. In 2025, there were clear regional differences in comments classified by Pangram as containing AI-assisted or -generated content, from 6.6% in Europe to 48.6% in Africa. In addition, we found that the increase in raw publication numbers was driven by comments classified by Pangram as containing AI-assisted and -generated writing. The detector had a specificity of 1.00 and a sensitivity of 0.98. Comments with detectable AI were cited less often and were less likely to be cited more than the median ( p < 0.05). Conclusions Comments classified by Pangram as containing AI-generated and AI-assisted content are becoming more common, but they could be less cited and impactful on policy. Journals should set clear rules for disclosure to protect the quality of scientific work.