Transformer-based language models for population-level public health communication: an evidence map of deployment, reach and policy gaps

Evidence map of transformer-based language models applied to population-level public health communication: 552 studies from six databases and grey literature (2019–2026), reported using PRISMA-ScR against a prospectively registered protocol. Version 8 (23 September 2026). Adds an independent blinded human re-charting of the 75-study validation holdout on all ten variables, and a missed-deployment audit: 215 records the charting instrument had placed at earlier stages were hand-charted blind across three samples, none was reclassified as deployed, bounding missed deployments at five and establishing the reported figure as a floor (at most nine of 552 studies, 1.6%). The headline deployment count is four studies (0.7%) following full-text verification of every study the instrument classified as deployed, revised down from six. The review is reframed throughout as covering transformer-based language models, with "large language model" reserved for generative models and the two architecture classes reported separately where the distinction bears on a finding. This deposit contains the preprint, its supplementary materials (Tables S1–S23) and the six main figures as standalone files. The full data, charting instruments, validation apparatus and deployment-audit worksheets are on the linked Open Science Framework project.