Operationalizing Governance for Youth-Facing LLM Well-being Support

Abstract

Youth are increasingly using large language models (LLMs) for well-being support, yet existing governance guidelines provide limited interaction-level requirements and empirical work rarely centers lived experience from youth and the communities that support them. We report a participatory study with 38 stakeholders (youth, parents, youth care workers). Using reflexive thematic analysis, we identify three stakeholder-derived governance domains for youth-facing LLM support: Interaction, Context, and Escalation. Stakeholders aligned on Governance of Interaction and Context but diverged on Escalation, particularly acceptable transitions to external support, with youth proposing intermediary peer referral mechanisms via personalized information retrieval. We translate these stakeholder accounts into preliminary normative requirements and dialogue inspection examples that can be used to guide model fine-tuning and post hoc evaluation. We further discuss how our formative findings expand upon major artificial intelligence guidelines by surfacing where existing frameworks remain underspecified for conversational governance and evaluation.