Mental Health AI Safety Claims Must Preserve Temporal Evidence

The paper introduces the concept of Temporal Safety Non-Identifiability, highlighting the inadequacy of current evaluation methods for mental health AI that overlook the temporal dynamics of interactions. It presents SCOPE (Safety Claims Over Preserved Evidence) and its specific application, SCOPE-MH, aimed at aligning safety claims with the evidence retained during evaluations. The authors demonstrate SCOPE-MH using the AnnoMI dataset, revealing critical failure mechanisms that traditional per-turn scoring fails to capture, emphasizing the necessity of preserving temporal evidence for safe deployment of mental health AI systems.