AI Chatbots for Mental Health Support (AI Risk Assessment)

Key Findings

• — Overall rating: 'Unacceptable Risk' — teens should not use general-purpose chatbots (ChatGPT, Claude, Gemini, Meta AI) for mental health support
• — Chatbots consistently failed to recognize warning signs of anxiety, depression, ADHD, OCD, PTSD, eating disorders, mania, and psychosis, despite improvements on explicit suicide/self-harm content
• — Safety performance degraded significantly in extended multi-turn conversations versus single-turn testing — the typical teen usage pattern
• — Systems are engineered for engagement: responses end with follow-up questions that prolong interaction rather than hand off to professional help
• — Chatbots lack the capabilities needed for safe mental-health support: clinical assessment, therapeutic relationship, coordinated care, and real-time crisis intervention

Methodology Notes

Qualitative red-team style risk assessment using teen test accounts (with teen protections enabled where available) across four major consumer chatbots; both single-turn prompts and extended conversations; clinical review via Stanford Brainstorm Lab. Not a quantitative benchmark — no published pass-rate statistics or fixed prompt set; product versions tested are a point-in-time snapshot. NB: sometimes dated to the Nov 20, 2025 press release; the assessment page itself is dated Nov 14, 2025 — mid-2026 'new report' framings trace back to this assessment.

Sources