Risk-Calibrated Patient-Facing AI Safety Cards: A UI/UX Benchmark for Explainable Medical AI Response Interfaces

Patient-facing medical AI systems communicate risk at moments when a non-expert user may act on what they read. This study evaluates risk-calibrated AI safety cards as a UI/UX framework for explainable medical response interfaces. The main experiment used 466 PatientSafetyBench prompts across five patient-safety categories: harmful medical advice, misdiagnosis and overconfidence, unlicensed practice of medicine, health misinformation, and bias or stigmatization. For each prompt, five response interfaces were generated: plain text, risk-label card, refusal-plus-explanation card, evidence-disclosure card, and next-step action card. The evaluation reports deterministic rubric-based communication scores rather than clinical safety outcomes. Across 2,330 PatientSafetyBench responses, the integrated Next-Step Action Card lowered the mean communication-risk score from 3.85 to 1.00 on a 1-5 scale, lowered the overconfidence-indicator score from 3.38 to 1.41, increased actionability from 16.46 to 98.09 on a 0-100 scale, increased risk-label clarity from 10.20 to 94.97, and increased evidence disclosure from 22.82 to 96.98. A second analysis used HealthBench physician-created rubric criteria to test whether the card structure aligned with communication, context, uncertainty, and escalation expectations in broader health conversations. The action-card condition increased communication-rubric coverage from 11.23% to 85.29% in HealthBench OSS, from 10.44% to 83.86% in HealthBench Consensus, and from 9.88% to 86.50% in HealthBench Hard. These results support the safety card as a reproducible information-design intervention for risk communication. They do not establish real-world patient behavior change or clinical safety; clinician review, patient testing, multilingual adaptation, and live-system evaluation remain necessary before deployment.