Mental Health AI Safety Claims Must Preserve Temporal Evidence

Current mental health AI evaluations may reach false safety conclusions because they typically assess individual responses or overall outcomes while ignoring how interactions unfold over time—meaning they miss clinically important failures like delayed crisis escalation, dependency formation, or gradual deterioration across multiple turns.

Temporal Safety Non-Identifiability is a formal problem: safety properties that depend on sequence, timing, or accumulation cannot be verified by evaluation methods that discard these temporal features, making it mathematically impossible to certify safety based on isolated response scoring alone.

The paper's proof-of-concept on motivational interviewing conversations demonstrated that per-turn behavior scoring fails to capture failure mechanisms that become visible when preserving the full sequence of interactions—suggesting current evaluation practices may be systematically missing real harms.

SCOPE-MH proposes a new reporting standard requiring mental health AI developers to align their safety claims with the actual evidence their evaluations preserve, meaning claims about safety should only be made about what the evaluation method can actually measure (e.g., if you only test individual responses, you cannot claim the system is safe over extended conversations).

Temporal evidence preservation is positioned as necessary, not optional for deploying mental health AI safely, implying that current industry practices may be insufficient and that evaluation infrastructure needs fundamental redesign before these systems should be widely used in clinical contexts.