Public Trust in Artificial Intelligence-Generated Health Information: A Critical Narrative Review of Health Literacy and Source Credibility

Abstract

Conversational systems built on large language models have become an ordinary route through which members of the public obtain health information, answering questions directly rather than returning ranked links to identifiable publishers. Whether such information is acted upon depends less on its accuracy than on whether it is trusted, and two explanatory traditions have been recruited to explain that trust: a receiver-centred tradition rooted in health literacy, and a message-centred tradition rooted in source credibility. This critical narrative review examines how well these traditions account for public trust in artificial intelligence-generated health information, where they conflict, and what the resulting evidence base can and cannot support. Literature was identified through Europe PMC, Crossref Metadata Search and OpenAlex, supplemented by citation searching and an authoritative institutional source, with a final search date of 9 June 2026. The synthesis is organised around conceptual foundations, observed patterns and determinants of trust, source-credibility cues and the attribution problem, health literacy as a contested moderator, the quality of the object of trust, and consequences for behaviour and equity. Three findings emerge with reasonable confidence. First, the widely assumed penalty attached to machine authorship is neither consistent nor robust: experimental and meta-analytic evidence points to small, conditional and sometimes absent effects, and blinded quality evaluations are routinely conflated with labelled trust experiments. Second, health literacy and its digital derivatives perform poorly as predictors of trust in representative surveys, whereas artificial intelligence literacy and perceived usefulness perform better, which suggests that the receiver-side account has been specified at the wrong level. Third, the stylistic properties that raise perceived credibility, particularly fluency and warmth, are partly antagonistic to accuracy, creating a structural rather than incidental risk. The evidence base remains dominated by cross-sectional designs, self-reported constructs and single-country convenience samples, and almost no work measures whether trust is calibrated to actual accuracy. Research priorities should shift from explaining trust levels to measuring trust calibration, and from self-reported literacy to performance-based appraisal.