Beyond Simulations: What 20,000 Real Conversations Reveal About Mental Health AI Safety

Abstract Large language models (LLMs) are increasingly used for mental health, yet safety evaluations rely primarily on small, simulation-based benchmarks removed from real-world language. We replicate four published safety evaluations assessing suicide risk handling, harmful content generation, and jailbreak resistance for general-purpose frontier models and a purpose-built mental health AI. We then conduct an ecological audit of 20,000 real user conversations with the purpose-built system, which includes layered safeguards for suicide and non-suicidal self-injury (NSSI). The purpose-built AI was significantly less likely than general-purpose LLMs to produce harmful content across suicide/NSSI (.4-11.27% vs 29.0-54.4%), eating disorder (8.4% vs 54.0%), and substance use (9.9% vs 45.0%) benchmarks. In real user data, clinician review found zero suicide-risk cases without crisis resources. Three NSSI mentions (.015%) lacked intervention, implying a .38% lower-bound false negative rate. Findings support the utility of ecological audits for safety estimation.