A new study in the Journal of Clinical Epidemiology puts numbers to what many researchers have suspected: the explosion of low-quality, formulaic manuscripts analyzing open-access health datasets may be systematic, AI-assisted, and accelerating.
Spick and colleagues analyzed publication trends to examine research derived from 36 major health and biomedical datasets indexed in the OpenAlex database between 2014 and 2025. They used a forecasting algorithm with three components - past data, trend and seasonality consideration, and noise removal - to model expected publication rates based on pre-2023 trends. The authors then measured how far actual publication counts deviated from those projections after the widespread adoption of generative AI tools in late 2022.
Nine datasets stood out as likely targets of exploitation: NHANES, UK Biobank, the FDA Adverse Event Reporting System (FAERS), FinnGen, the Global Burden of Disease Study, MIMIC III/IV, the China Health and Retirement Longitudinal Study (CHARLS), CDC WONDER, and TriNetX. Together, these nine datasets had an estimated excess of roughly 11,577 publications above predicted trends in 2025 alone, a three-fold increase over their 2022 combined publication count.
Additionally, the authors found warning signs that point to AI use. Titles were becoming strikingly formulaic. For articles utilizing the CDC WONDER database, the word "trends" appeared in 69% of titles in 2025, up from 20% in 2022. For FAERS, the acronym itself appeared in the title of 55% of publications, up from 26%.
Beyond quality concerns, papers following templates like "The association of [Predictor A] with [Outcome B] in [Population C] using [Open Access Dataset D]" introduce false discoveries into the literature (i.e., present statistically significant findings from large datasets without a corrected p-value), dilute genuine findings, and erode confidence in otherwise valuable data resources. For example, FAERS is a meaningful tool for drug safety monitoring, but only if the signal is not buried under thousands of (potentially) poorly conducted association analyses. The problem extends to AI language models themselves, which are trained on published literature: corrupted inputs will propagate corrupted outputs.
This matters enormously for systematic review teams. When a body of literature is flooded with papers sharing the same template, the same analytical shortcuts, and potentially the same selectively reported findings, the evidence base being synthesized becomes harder to trust. These papers are appearing in indexed journals and clearing peer review, meaning your literature search will likely pull these articles alongside other non-AI written articles. The authors estimate that excess publications represented roughly 50% of 2025 output for the nine most affected datasets.
The practical response is a renewed commitment to critical appraisal at every stage of the review process.
Systematic reviewers should consider the plausibility of findings, the transparency of methods, and whether outcomes appear to have been pre-specified. Formulaic titles, an absence of pre-registration, implausible effect sizes, and a lack of false discovery correction are all worth treating as warning signs during screening and full-text review.
Searches that pull from datasets like NHANES, UK Biobank, or the Global Burden of Disease Study are now operating in a considerably noisier environment, and careful critical appraisal, with attention to formulaic methodology, implausible findings, and selective outcome reporting, is more important than ever.
References
Spick M, Onoja A, Harrison C, Stender S, Byrne J, Geifman N. Quantifying new threats to health and biomedical literature integrity from rapidly scaled publications and problematic research. J Clin Epidemiol. 2026 May;193:112203. doi: 10.1016/j.jclinepi.2026.112203. Epub 2026 Feb 23. PMID: 41740900; PMCID: PMC13178217.
https://www.jclinepi.com/article/S0895-4356(26)00078-8/fulltext
