The teen chatbot safety problem is a product problem, because 63% told nobody

A group of clinical psychologists and a child psychiatrist, writing at phys.org, report that 72% of teens have interacted with AI companion chatbots, and that in a separate study 63% of the teens and young adults who used an AI chatbot for mental health advice had not told anyone they did ^[1][3]. Between 13% and 20% sought support through AI for emotional concerns during a two-month window in 2025 ^[2]. If most help-seeking is invisible to adults, then the safety controls that matter are the ones shipped in the product, not the ones a parent is asked to enforce.

Arithmetically, no more than about 37% of that help-seeking cohort disclosed it to anyone ^[1]. Non-disclosure is the majority case, which means any intervention gated on a conversation between a teen and a trusted adult starts from behind ^[2].

The authors are blunt about what teens are actually using. Most are on general-purpose tools such as ChatGPT, Claude and Gemini, or assistants embedded in apps they already have, such as Snapchat's My AI, and those systems were never designed or validated for mental health support ^[4][5][6]. The evidence base people cite belongs to a different product category: Woebot, a fully automated CBT-based chatbot, reduced depressive symptoms in college students after two weeks, and there is evidence some therapeutic chatbots help adults with mild symptoms when used alongside professional care ^[7][8]. That is a claim about a supervised adjunct, not about a general assistant at 2 a.m. A November 2025 Common Sense Media report rated general-purpose chatbots as unacceptably risky for teen mental health support ^[9].

The specific design failures the authors name are the kind that show up in a spec review. General-purpose chatbots typically lack effective crisis protocols or pathways to escalate care when someone is in danger ^[10]. Studies of both therapeutic and general-purpose chatbots found inadequate responses to thoughts of suicide, and even when a bot surfaces the 988 Suicide and Crisis Lifeline, conversations with concerning content are often allowed to continue with no hard stop ^[11]. Safety of the interactions degrades over the course of long conversations ^[12]. Surfacing a resource and then continuing the session is a disclosure pattern, not a safety mechanism; it satisfies an audit line while leaving the risky path open.

The rest compounds it. These systems can generate inaccurate or misleading advice delivered with high confidence, which can reinforce unhealthy beliefs ^[13]. Teens sometimes describe chatbot conversations as emotionally meaningful and occasionally prefer them to talking with friends, and adolescents experiencing loneliness or anxiety may form strong attachments that displace human connection and distort expectations of real relationships ^[14][15]. Products optimised to simulate empathy and maintain engagement are, by construction, pulling in the opposite direction from a handoff ^[16].

Worth noting where the clinicians themselves have landed: having published guidelines for clinicians in 2025 on addressing teen chatbot use safety, they are now working on the same for parents and other trusted adults ^[17]. That sequence is an admission that the clinical encounter was not catching this either.

What to watch: whether vendors publish safety performance as a function of session length rather than single-turn benchmarks, given that degradation over long conversations is the documented failure mode ^[12]; whether crisis handling moves from resource text to enforced interruption, since continuation without a hard stop is what the studies found ^[11]; and whether any parental control shipped this year works without requiring the teen to disclose, because 63% did not ^[3].