JMIR Publications

ABSTRACT

Large language model–based virtual patients have expanded access to conversational practice, individualized feedback, and repeatable clinical cases. However, clinical competence is enacted across people, records, diagnostic systems, time pressure, handovers, competing priorities, and institutional constraints—not only during a learner–patient dialogue. This Viewpoint proposes the AI-native virtual hospital as an evidence-informed maturity model for medical education. An AI-native virtual hospital is a simulated clinical ecosystem designed from the outset around longitudinal patient states, multimodal clinical information, role-specific agents, workflow constraints, consequences of decisions, educational orchestration, and human-supervised governance. The model distinguishes 3 levels: conversational fidelity in virtual patients; workflow and data fidelity in virtual clinical environments; and team, system, and ecosystem fidelity in AI-native virtual hospitals. Potential educational uses include interprofessional teamwork, dynamic clinical reasoning, documentation, escalation, longitudinal care, and repeated deliberate practice. Yet greater technical complexity does not guarantee greater educational value. These environments may create convincing but invalid experiences, compound errors across agents, increase cognitive load, reproduce inequity, and expose sensitive learner data. Implementation should therefore begin with bounded formative use cases, expert-authored case logic, transparent agent roles, versioned models, human debriefing, and prospective validity and fairness studies. AI-native virtual hospitals should complement—not replace—standardized patients, simulation centers, faculty supervision, real patients, and workplace-based learning.