Mental Health AI Safety Claims Must Preserve Temporal Evidence

A recent study published on arXiv highlights significant flaws in the current assessments of AI safety in mental health, primarily due to their focus on inappropriate temporal scales. The researchers assert that failures with clinical significance—like delayed escalation, dependency formation, and gradual deterioration—are influenced by the sequence and accumulation of interactions rather than merely isolated responses. They propose the concept of Temporal Safety Non-Identifiability, which clarifies why safety properties reliant on timing and sequence cannot be validated through protocols that overlook these aspects. To address this, they introduce SCOPE (Safety Claims Over Preserved Evidence), a framework designed to align safety assertions with the evidence retained in evaluations. The paper can be found on arXiv under ID 2605.08827.