Media amplification, model source cues, and expectancy violation in public acceptance of generative AI: evidence from a health-consultation experiment.

BACKGROUND: Generative artificial intelligence (GenAI) is increasingly used for health-information seeking, where evaluations may depend on prior media amplification, the model identity presented, and expectancy violation. METHODS: We conducted a 2 (media amplification: benefit vs. risk) × 2 (model source: general-purpose vs. professional) × 2 (expectancy violation: positive vs. negative) between-subjects online experiment with 491 participants. Participants read prior-user reports of better- or worse-than-expected performance without using the AI or observing a response. Type III factorial models tested attitude toward use (ATT), behavioral intention (BI), and perceived risk (PRISK). A secondary analysis estimated conditional indirect associations between expectancy violation and BI through PRISK. RESULTS: Media amplification increased PRISK but did not change ATT or BI on average. Professional model-source cues and positive expectancy violation produced favorable average effects across all three outcomes. In the primary Type III models, the three-way interactions reached significance for ATT, BI, and PRISK (partial η2 = 0.008-0.009). Sensitivity support was strongest for the PRISK interaction; the ATT and BI terms varied across specifications. Conditional indirect associations between expectancy violation and BI through lower PRISK were observed in three of the four conditions, while the joint-moderation index included zero. CONCLUSION: In this health-consultation vignette, media amplification most clearly altered perceived risk, while model source and expectancy violation shaped prospective acceptance across communication conditions. The three-way evidence was most consistent for perceived risk.