Mental Health AI Safety Claims Must Preserve Temporal Evidence

A new paper argues that current safety evaluations for mental health AI are structurally incapable of detecting clinically dangerous interaction patterns because they discard the sequence and timing of conversations [1]. The authors introduce a formal principle called SCOPE to align safety claims with the evidence an evaluation actually preserves [2]. The work, submitted to arXiv on 9 May 2026 and revised on 16 June 2026, identifies what the authors term Temporal Safety Non-Identifiability — a formal account of why safety properties that depend on sequence, timing, accumulation, or recovery cannot be certified by protocols that discard those features [2]. Ratna Kandala is listed as the corresponding author on the submission [1]. Current evaluation practice typically scores isolated responses, endpoint outcomes, or aggregate dialogue quality [2]. The paper contends that clinically consequential failures — including delayed escalation, repeated reinforcement, dependency formation, failed repair, and gradual deterioration across turns — may arise from the order and accumulation of interactions themselves [2]. Because standard protocols do not retain temporal evidence, they can produce invalid safety conclusions [1]. Mental health, as defined by the World Health Organization, encompasses emotional, psychological, and social well-being and affects how people manage stress and make everyday decisions [4]. Deploying large language models in this domain raises distinct risks. LLMs are neural networks trained on vast text corpora for natural language processing tasks, and biased or inaccurate training data can make their outputs less reliable [3]. The field of artificial intelligence has seen an acceleration in generative capabilities since the 2020s, accompanied by growing attention to AI safety and unintended harms [5]. From the formalization, the authors develop SCOPE — Safety Claims Over Preserved Evidence — as a general principle, and instantiate it as SCOPE-MH, a mental-health reporting standard [2]. They operationalize SCOPE-MH through a proof-of-concept on the AnnoMI dataset of expert-annotated motivational interviewing conversations [1]. The analysis reveals mechanisms of failure that per-turn behavior scoring does not represent [2]. The paper proposes SCOPE-MH as a diagnostic complement to existing evaluation infrastructure and argues that evaluation preserving temporal evidence is necessary, not optional, for safety-critical mental health AI deployment [2]. The submission totals 853 KB in its revised form [1].