Currently submitted to: JMIR Infodemiology
Date Submitted: Nov 8, 2026
Open Peer Review Period: Nov 17, 2026 - Jan 12, 2027
(currently open for review)
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Drug Safety Warnings in English, Russian, Arabic and Hebrew Wikipedia: A Cross-Sectional Audit of 33 Drugs Across Three Risk Domains
ABSTRACT
Background:
Wikipedia is a leading consumer health resource and a heavily weighted source for the large language models and AI search tools that increasingly answer health questions, yet its medical content is audited almost only in English. Whether patient-actionable drug-safety warnings are equally present in other language editions is unknown.
Objective:
We tested whether Russian, Arabic, and Hebrew Wikipedia articles carry less actionable drug-safety warnings than English articles, and whether any gap recurs across unrelated safety domains.
Methods:
In this cross-sectional audit we examined 33 drugs in four language editions: 24 risk-positive drugs (eight each for glucose-6-phosphate dehydrogenase [G6PD] deficiency, teratogenicity, and QT-interval prolongation, benchmarked to the Clinical Pharmacogenetics Implementation Consortium [CPIC] guideline, US Food and Drug Administration [FDA] pregnancy labelling, and the CredibleMeds Known-Risk list) and nine negative controls. Three large language models, used as a measurement instrument, coded each warning 0 (not mentioned), 1 (mentioned), or 2 (explicit directive); absent articles were scored 0. A pharmacist resolved disagreements for G6PD and teratogenicity; the 12 QT disagreements kept the majority code, with an extreme-code sensitivity analysis. The prespecified analysis was a hierarchical Bayesian cumulative-logit model with English as reference. Exploratory analyses examined article coverage, length, and escalation from mention to directive. We also audited Wikidata and pageviews.
Results:
Articles existed for all 33 drugs in English, 32 in Arabic, 27 in Hebrew, and 26 in Russian, and drew 10,897,920 views from July 2025 to June 2026. Pooled cumulative odds ratios versus English were 0.41 (Russian), 0.25 (Arabic), and 0.31 (Hebrew), with posterior probabilities of 78%, 87%, and 87% that English warns more strongly; all three 95% credible intervals included 1. Eight of nine language-by-domain estimates were below 1; the exception was Russian QT (odds ratio 1.06), and the largest gap was G6PD in Russian (odds ratio 0.05). In exploratory analyses, restricting the model to existing articles moved the Russian estimate to 0.95, suggesting a coverage gap, whereas Arabic articles that mentioned a risk escalated it to an explicit directive in 7 of 15 cells against 19 of 23 in English (odds ratio 0.22, 95% credible interval 0.02 to 0.73), a length-independent escalation gap. Wikidata's contraindication property appears in only 12 statements across the knowledge base, none for the audited drugs.
Conclusions:
Drug-safety warnings were weaker in Russian, Arabic, and Hebrew Wikipedia than in English in eight of nine comparisons across three unrelated classes of drug harm, with wide uncertainty from a small purposive panel and model codes not yet validated against human coding. Because Wikipedia feeds the AI systems that patients consult, these gaps are candidate sources of downstream health information disparities; populating Wikidata's contraindication property is a concrete, low-cost place to intervene.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.