Trust Calibration and Epistemic Uncertainty Markers across Medical Advice Assistants

Abstract

The rapid integration of artificial intelligence into healthcare has introduced novel paradigms for patient interaction, most notably through Medical Advice Assistants. While these systems offer unprecedented access to health information, they concurrently present profound risks associated with algorithmic overconfidence and user automation bias. This paper investigates the intricate relationship between the linguistic expression of epistemic uncertainty by medical artificial intelligence systems and the subsequent calibration of user trust. We employ a novel graph analysis framework to model the cognitive and conversational alignment between human users and conversational agents. By representing dialogic interactions as semantic networks, we analyze topological features such as modularity, network density, and centrality distributions. The study involves a controlled experiment wherein participants interacted with medical advice systems exhibiting varying degrees of epistemic uncertainty markers. Our findings indicate that the strategic insertion of these markers significantly alters the semantic graph structure of the conversation, shifting users from passive acceptance to active, analytical engagement. This shift is quantitatively linked to optimal trust calibration, mitigating both detrimental over-trust and unwarranted under-trust. The results offer critical design implications for the development of safe, transparent, and user-centric medical artificial intelligence systems.