JMIR Publications

ABSTRACT

Recent studies have shown that consumer-facing artificial intelligence chatbots can produce medical information that is unreliable, overconfident, and poorly grounded in evidence. These findings are important, but they do not justify treating all AI used in healthcare as though it poses the same kind of risk. General-purpose consumer chatbots, clinician-integrated decision support tools, and autonomous clinical AI differ in their intended users, knowledge boundaries, governance arrangements, and acceptable failure modes. Yet current debate still often collapses them into a single category of "medical AI," making both regulation and procurement less precise. This article proposes a three-category design taxonomy for clinical AI to distinguish these systems more clearly and to focus attention on the features that matter most in practice: bounded evidence, transparent grounding, human oversight, and the ability to abstain when a system should not answer. The aim is not to claim that any category is inherently safe, but to offer a practical framework for clinical governance, evaluation, and safer deployment.