Risk-Calibrated Patient-Facing AI Safety Cards: A UI/UX Design Framework for Rubric-Based Medical Risk Communication

Patient-facing medical AI systems can provide valuable health information; however, safety-sensitive queries require responses that establish clear boundaries while remaining informative, respectful, and actionable. This paper presents the Risk-Calibrated Safety Card, a UI/UX framework for communicating medical AI responses in high-risk situations. The framework transforms safety-sensitive outputs into a structured card containing five elements: risk level, explanation of why an unrestricted response may be unsafe, bounded safe information, professional-help guidance, and a bias-sensitive language note. The evaluation uses HealthBench, a benchmark of realistic health conversations with physician-authored rubrics, including the HealthBench Full evaluation split and robustness analyses on the Consensus and Hard subsets. Four response formats were compared: unstructured answer, refusal-only answer, refusal with explanation, and the proposed safety card. Across 4,597 HealthBench Full records, the safety card achieved the lowest rubric-based safety-communication risk score (1.27), the highest weighted positive-rubric coverage (0.664), and complete coverage of predefined card components (1.00). Refusal-only responses reduced unsafe personalization but showed limited helpfulness (1.55) and negligible positive-rubric coverage (0.001). Refusal with explanation improved boundary communication but lacked the structured presentation provided by the card. Although the safety card produced longer responses and a slightly lower readability score than the refusal-with-explanation condition (Flesch-Kincaid grade 14.39 vs. 14.72), the results suggest that structured safety cards can improve the visibility of risk, guidance, and support cues. These findings represent rubric-based interface evidence and should not be interpreted as validation of patient outcomes, clinical safety, or real-world deployment effectiveness.