The interactive turn in generative AI for self-harm

Key Findings

• — Self-harm is functionally heterogeneous (affect regulation, self-punishment, interpersonal communication), so a classification output supports risk stratification but does not explain what the behaviour does for the person; the authors argue clinical intervention depends on that functional understanding
• — The proposed shift is described as logical rather than technological: from one-shot classification toward iterative, supervised, safety-bounded functional assessment in dialogue
• — The paper names domain-specific interaction risks: an insufficiently guarded interactive AI could function like pro-self-harm online communities (social contagion, normalisation, method sharing) if it fails to recognise when its responses are read as permission or encouragement
• — Mitigations proposed go beyond content filters: guardrails that interrupt validation-seeking patterns, protocols for shifting from functional exploration to alternative generation, hard boundaries around method discussion, and regular active encouragement of human help-seeking
• — The outlined architecture includes trajectory-based safety assessment across a conversation, escalation pathways to human care, participation and co-production with people with lived experience, and human oversight; the authors state that none of the proposed mitigations has been validated for self-harm populations
• — The authors state that current evidence from conversational AI and digital mental health supports the plausibility of such interaction but does not establish its safety, effectiveness or clinical utility for self-harm or NSSI populations

Methodology Notes

Conceptual and theoretical paper (Frontiers in Psychiatry, volume 17, article 1877825; received 11 May 2026, accepted 31 July 2026, published 25 August 2026; CC BY 4.0). No empirical data; the authors distinguish evidence-supported claims from hypotheses requiring testing and label the proposal a research agenda, not a validated clinical intervention. Funding partly from Zhejiang provincial sources per the funding statement; the authors declare no conflict of interest and declare generative-AI use in preparation. Verified via PubMed (PMID 42713193, EDAT 2026-09-09) and Crossref; full text read from the Europe PMC deposit (PMC13551726).

Sources