The interactive turn in generative AI for self-harm

Self-harm, particularly non-suicidal self-injury (NSSI), presents a distinctive challenge for artificial intelligence in mental health. The dominant paradigm frames AI as a classifier: ingest multimodal signals, output a risk label. This paper argues that this approach faces a principled limitation. Self-harm is functionally heterogeneous; the same behavior can serve different purposes across individuals and contexts, from affect regulation to self-punishment to interpersonal communication. A classification output can support risk stratification, but it does not by itself explain what the behavior is doing for the person, and clinical intervention often depends on that functional understanding. Generative AI opens a possible alternative path. Rather than treating AI only as a detector that reads signals and returns a verdict, AI can be conceived as an interactive system capable of sustained, function-oriented dialogue. Current evidence from conversational AI and digital mental health supports the plausibility of such interaction, but it does not establish the safety, effectiveness, or clinical utility of interactive AI for self-harm or NSSI populations. This article therefore presents the interactive turn as a conceptual and empirical research agenda, not as a validated clinical intervention. Multimodal signals may eventually enrich this dialogue, but the core shift is logical rather than technological: from one-shot classification toward iterative, supervised, safety-bounded functional assessment. The paper develops the theoretical case for this shift, distinguishes evidence-supported claims from hypotheses that require testing, and outlines a system architecture that combines a conversational agent, cognitive and safety layers, trajectory-based monitoring, escalation pathways, multimodal inputs, and human oversight.