From Explanation to Persuasion: A Systematic Review of the Effects of AI-Generated Health Materials
Jiaxiang Xu, Jingyuan ShiIntroduction:
Artificial intelligence (AI), particularly large language models (LLMs), is increasingly used to generate audience-facing health materials. However, evidence regarding how such materials influence human audiences remains fragmented. This systematic review aims to synthesize experimental evidence on the effects of AI-generated health materials across different communication functions and outcome domains.
Methods:
A systematic search was conducted across Web of Science, PubMed, Scopus, and PsycINFO. Eligible studies were required to (1) involve health-related contexts, (2) use AI-generated materials as the primary exposure or intervention, (3) include human participants, and (4) assess post-exposure outcomes. Sixteen studies met the inclusion criteria. Due to substantial heterogeneity in study designs, outcome measures, and intervention formats, a narrative synthesis approach was adopted. Outcomes were categorized into Content Evaluations and Communicative Quality, Comprehension and Knowledge Outcomes, Affective Outcomes, Attitudinal and Belief-Related Outcomes, Behavioral and Adherence Outcomes, and key boundary conditions (e.g., authorship disclosure, tailoring strategy, and communication context) were examined.
Results:
AI-generated health materials demonstrated the most consistent benefits in explanation and simplification functions, with uniformly positive findings in comprehension outcomes and several cases of superiority or noninferiority compared to human-generated materials. Evaluative and persuasion-related outcomes were more heterogeneous and were often contingent on factors such as AI authorship disclosure and contextual cues. Persuasive effects were inconsistent and did not reliably translate into behavioral intentions. Behavioral outcomes were most evident when AI-generated materials reduced informational barriers, suggesting a comprehension-mediated pathway to action.
Conclusions:
Current evidence indicates that AI-generated health materials function more reliably as tools for explanation and simplification than for persuasion. Their effectiveness in influencing attitudes and behavior is conditional and context dependent. Future research should further examine the mechanisms and boundary conditions shaping these effects, particularly in relation to authorship disclosure and targeting strategies.