<![CDATA[Background: AI-generated health advice is increasingly encountered by Indonesian users as a convenient source of explanation, reassurance, and self-care guidance, yet its linguistic fluency may conceal biomedical incompleteness, weak uncertainty marking, and unsafe action cues. Objective: This study aims to evaluate Indonesian AI-generated health responses by examining how accuracy, uncertainty communication, and communicative risk interact across safety-sensitive health domains. Method: Using a qualitative-dominant corpus design, this study analysed 15 Indonesian prompt-response units linked to official public-health references and coded each response for accuracy, completeness, disclaimer relevance, referral specificity, directive force, tone, and risk level. Results: Findings show that only 4 responses were both accurate and complete, whereas 5 were accurate but incomplete, 4 were partially aligned with weak red-flag articulation, and 2 were misleading or unsafe. Uncertainty was often present as generic disclaimer language, but it was not consistently translated into specific referral advice or usable escalation thresholds. Implication: Communicative risk emerged when empathetic, calm, or accessible responses softened urgency, normalised self-management, or implied authority beyond available clinical information. Novelty: This study contributes a linguistic-pragmatic model for AI-health evaluation by demonstrating that safe advice requires alignment between factual accuracy, explicit uncertainty, and responsible communicative force.]]>
Copyrights © 2026
