From simulation to pedagogy: structured AI standardized patients for clinical communication training validated through multi-model and randomized evaluation

Standardized patients (SPs) are central to clinical communication training but are constrained by cost, scalability, and reliance on trained actors. We present AI standardized patients (AI-SPs), large language model-driven simulators governed by a three-layer information architecture that modulates disclosure according to learner skill. We validate this approach across three phases. In Phase 1, blinded expert evaluation of 350 simulated consultations from five frontier LLMs showed that learner skill level, rather than model choice, drove performance variation (2 = 0.31 vs 0.06), indicating that pedagogical quality emerges from architec-tural design rather than model scaling. In Phase 1b, 155 live student consultations revealed systematic failures in eliciting safety-critical information, generating automated curriculum diagnostics without expert observation. In a three-arm pilot randomized controlled trial (Phase 2, n = 58), AI-SP training achieved skill gains non-inferior to human SP practice, with a distinctive self-efficacy benefit unique to the AI-SP arm. These findings suggest that architecture-driven AI-SPs offer a scalable, model-portable paradigm for clinical communication training.