Generative AI Research in Infectious Disease Skews Toward High-Income Conditions and Under-Reports Safety: A 524-Study Scoping Review

Large language models (LLMs) are being rapidly applied to infectious disease, yet no comprehensive assessment examines whether this research activity addresses the diseases causing greatest global morbidity or whether safety has been evaluated. This scoping review, following JBI methodology and PRISMA-ScR guidelines with a pre-registered protocol, searched PubMed, Embase, Scopus, Web of Science, supplementary sources, and a targeted post-hoc search for neglected tropical diseases (NTDs) for studies evaluating LLMs or generative AI for any infectious disease task from November 2022 to April 2026; two studies from the NTD search published in May 2026 were retained as a documented protocol deviation. From 42,030 records in the main searches, 61 from the NTD search and 63 from a corrected forward citation-chaining run, 524 studies were included. Research was concentrated on COVID-19 (25.4%) and conditions prevalent in high-income countries. Among approximately 20 WHO-prioritised NTDs, only scabies (2 studies) and leishmaniasis (3 studies) were the subject of single-disease studies, and one multi-NTD diagnostic-accuracy study covered Chagas disease and visceral leishmaniasis. Schistosomiasis, lymphatic filariasis, onchocerciasis, soil-transmitted helminthiases (which alone infect an estimated 1.5 billion people), and African trypanosomiasis, all concentrated in low- and middle-income countries (LMICs), had zero LLM-focused studies even on targeted searching. Only 12 studies (2.3%) addressed parasitic infections of any kind. Simultaneously, 510 studies (97.3%) did not report safety or hallucination assessment in their abstracts, including 8 of 10 clinically deployed applications; a full-text validation subsample (n=75) provided a calibration estimate for under-reporting. Only 10 studies (1.9%) reported clinical deployment. The generative AI evidence base for infectious disease parallels established patterns of medical innovation flowing toward commercially viable conditions in well-resourced settings, with safety assessment rarely visible in published reporting, a combination that warrants attention before clinical deployment scales further. Review registration: https://doi.org/10.17605/OSF.IO/QE629 Data and code: https://osf.io/xbtne What changed since version 2: The corpus is expanded from 516 to 522 studies by a post-hoc NTD-targeted supplementary search (6 additional studies: 2 on scabies, 3 on leishmaniasis alone, and 1 multi-NTD diagnostic-accuracy study covering Chagas disease and visceral leishmaniasis). The version 2 statement that no included study addressed neglected tropical diseases is superseded. Two of these studies were published in May 2026, after the April 2026 eligibility window, and are retained as a documented protocol amendment (amendment log on the OSF project page). All corpus counts, Table 1, Table S1 and every figure are recomputed from a consolidated 524-study extraction, in which the NTD-search and citation-chaining studies are coded with the same abstract-level classifier. Abstract-level non-reporting of safety or hallucination assessment is 510/524 (97.3%), compared with 503/516 (97.5%) in version 2. Forward citation chaining is corrected. The original run had chained two records other than the stated seed reviews; none of its records was included. A re-run on the correct seed records (records published to 6 April 2026) screened 29 new records and added 2 studies, giving 524. Backward citation chaining, specified in the protocol, was not performed and is now declared as a deviation (Amendment 4). A stratified full-text validation subsample (n=75) is added to calibrate the abstract-level safety-reporting estimate. It was drawn from the 516 studies included before the NTD-targeted search; the studies added by that search and by the citation-chaining re-run were not in its sampling frame (Supplementary Appendix F). The title, abstract and discussion are reframed around disease equity and safety reporting; the limitations are expanded; the description of the medRxiv/bioRxiv search and of the model landscape is corrected; every reference was re-verified against CrossRef, PubMed, DataCite or the source page, and DOIs were added. Supplementary figures are renumbered Figures S1 to S4, and the supplementary table Table S1. Licence changed from CC BY 4.0 (version 2) to CC BY-NC-ND 4.0.