Large language models for population-level public health communication: a computational evidence map of global reach, equity and deployment gaps

Preprint version 3 (posted 4 June 2026). Version 1 was posted on 6 April 2026 and version 2 on 3 June 2026. This version applies minor revisions arising from external review: a mis-rendered author name in one reference was corrected, a task × domain gap count was clarified (the denominator excludes the cross-cutting "General" domain), a closely related 2025 scoping review was cited for positioning, and the reference list was held at 60. No findings, analyses or conclusions changed. Large language models (LLMs) are increasingly proposed for population-level public health communication, yet whether their development reaches the populations most exposed to health misinformation is unknown. We conducted an automated, computationally assisted evidence map (reported using the PRISMA extension for scoping reviews for transparency), screening 30 715 records from six databases and grey literature (2019–2026). AI-assisted screening showed high agreement with a second screen (Cohen's κ=0.83), and rule-based charting reliability varied (74% overall, ~50% for interpretive variables such as ethics and evaluation). The field grew from 3 studies in 2020 to 215 in 2025 but is skewed away from global need. English dominated as a target language (83%); among the 474 studies with identifiable affiliations, first authorship concentrated in the United States (38%) and other high-income countries, while low- and middle-income countries (LMICs) and major languages such as Russian and French were almost absent. This direction holds, though exact magnitudes are entangled with the review's English-language scope. Misinformation detection (186 studies) exceeded correction (11) by over an order of magnitude, and fewer than 10% of tools were deployed. Realising the promise of LLMs for global public health requires deliberate investment in multilingual, LMIC-relevant, correction-oriented and deployed applications. The complete charting dataset (552 studies), screening decisions, validation sample and all analysis scripts are openly available on the Open Science Framework: https://doi.org/10.17605/OSF.IO/X5N78