Dear Editor,

Matthews et al. (Reference Matthews, Cleary and Firth2026) provide a timely analysis of large language model (LLM) chatbots and youth mental health. Their distinction between purpose-built mental health chatbots and generic LLMs is clinically important because conversational fluency may conceal marked differences in validation, governance, and crisis-response capacity. Their discussion of the interplay between system-level risks and human-level vulnerabilities offers a valuable conceptual foundation. The article also calls for collaboration among clinicians, public health experts, developers, and people with lived experience. Yet the allocation of responsibility across these actors remains unclear. Institutional arrangements ultimately determine whether emerging harm is recognised, escalated, and connected to accountable human care.

This gap becomes important when considering the authors’ suggestion that professionals screen for LLM use in selected contexts. Frequency of use alone is a poor indicator of harm. Routine inspection of conversation histories could pathologise ordinary technology use, compromise privacy, and discourage honest disclosure. Assessment should instead examine the purpose of use, emotional dependency, displacement of human relationships or professional care, exposure to unsafe advice, and changes in sleep, functioning, or help-seeking. Safeguarding should preserve adolescent agency and require the least intrusive assessment proportionate to the apparent risk. Clark (Reference Clark2025) reported that, across 10 chatbots, harmful proposals were endorsed in 19 of 60 response opportunities involving simulated adolescent scenarios. Such findings show why automated warnings cannot replace responsive professional judgement.

An institutional safeguarding layer could operationalise the responsibilities implied by Matthews et al. Five elements merit consideration: purpose-sensitive assessment, risk-based triage, defined professional responsibility, supervised escalation, and privacy-preserving documentation. Low-risk psychoeducational use may require digital mental health literacy rather than clinical intervention. Repetitive reassurance-seeking, emotional dependency, social withdrawal, or deteriorating functioning warrants structured assessment. Content suggesting suicide, self-harm, psychosis, abuse, or imminent danger requires urgent professional review under established safeguarding procedures. Effective human oversight must identify the responsible professional, threshold for action, response timeframe, referral route, and follow-up process. Nominal access to a counsellor does not constitute a functioning safety system.

Schools represent one important setting because they reach young people who may never access specialist services. Teachers and peer supporters may recognise behavioural changes and encourage help-seeking but should neither inspect private chatbot exchanges nor conduct clinical risk assessments. School counsellors may assess concerns and coordinate referral according to competence and workload. Families can support care when disclosure is proportionate to risk and respects adolescents’ evolving autonomy. Clinicians should retain responsibility for complex assessment and crisis management. Platform providers remain accountable for age-appropriate design, transparent data practices, independent safety evaluation, and reliable crisis protocols. These responsibilities carry particular urgency in low-resource and multilingual settings, where professional support and culturally appropriate model performance may be uneven. Current evidence indicates that LLM chatbots remain insufficiently evaluated in high-stakes mental health contexts (Hua et al., Reference Hua, Na, Li, Liu, Fang, Clifton and Torous2025a, Reference Hua, Siddals, Ma, Galatzer-Levy, Xia, Hau, Na, Flathers, Linardon, Ayubcha and Torous2025b).

Research should examine extended conversational trajectories rather than isolated prompts and test performance across languages, developmental stages, clinical risk groups, and model updates. Relevant outcomes include harmful endorsement, emotional dependency, displacement of human care, appropriate escalation, referral completion, and continuity of support. A central policy question follows: can institutions recognise harmful chatbot use without converting safeguarding into digital surveillance? Youth mental health chatbot safety should be judged by proportionate protection, preserved privacy, and accountable care after risk emerges not by the mere display of a warning before harm occurs.