A Developmentally-Grounded Framework for Evaluating LLM Behavior in Adolescent Depression and Suicide-Risk Conversations

The use of AI chatbots for mental health support is rising rapidly among young people, often outside the awareness of parents, clinicians, or other trusted adults [McBain et al., 2026]. Evaluating these systems is especially difficult in the context of adolescent depression and suicide, where risk is developmentally and socially embedded and real-world crisis conversations involving minors are difficult to collect, share, and benchmark safely. We introduce a clinician-led methodology for constructing and evaluating developmentally grounded, human–AI interactions, instantiated in 180 human-authored multi-turn conversations (20–30 talk turns) across six frontier LLMs simulating adolescents ages 13–17. Transitionally-aged youth role-players were trained to construct realistic personas and portray characteristics of depression and suicide-related risk across four pre- specified tiers. Conversations established a coherent persona and age before tar- geted risk emerged and underwent structured paired review for developmental plausibility, age signaling, persona consistency, intended risk level, and clinically relevant risk characteristics. We also describe a clinician-led annotation frame- work informed by Interpersonal Theory of Suicide [Van Orden et al., 2010] and the Columbia-Suicide Severity Rating Scale [Posner et al., 2011], with specific attention to developmentally specific AI safety concerns, including contextual risk and protective factors, relational and anthropomorphic behavior, and severity criteria informed by adolescent decision-making and impulsivity. The methodology provides a reproducible approach for evaluating high