Mixed reality and generative AI are opening new possibilities for clinical training and patient support. But expanding their role requires answering a harder question: What should we entrust to these systems, and what evidence should justify that trust? Drawing on our research, this keynote examines that question across three settings: clinical training, support beyond the clinic, and AI interactions without clinician oversight. In training, mixed-reality simulations move beyond fixed scenarios, allowing simulated patients’ conditions to evolve as trainees assess and act. AI-powered standardized patients extend rehearsal to clinical communication. For patient support, we explore wearables developed with rehabilitation clinicians and tools that help patients understand medical procedures. We also examine failures that matter when people turn to AI without a clinician present. In a workshop involving seventy clinicians, educators, and researchers, subtle changes in wording shifted models’ sensitivity to risk, while models sometimes agreed when they should have challenged a premise. Together, these projects highlight both opportunities to extend human expertise and the need to test the limits of that extension. The keynote considers what evidence is needed to move from promising demonstrations to justified roles in care. The goal is healthcare in which rehearsal comes before the stakes are real, patients understand what is about to happen to them, and systems are entrusted only with responsibilities they have demonstrated they can fulfil.