OpenAI Introduces MentalHealthBench for Realistic Mental Health Conversations

OpenAI says most AI evaluations in mental health have concentrated on emergency scenarios and measured success with broad, predefined criteria. [1] That emphasis has left less visibility into performance across the full range of mental health conversations and alignment with situation-specific expert guidance, beyond avoiding disallowed responses. [1] OpenAI says evaluating these varied situations is essential to building AI that actively supports people’s long-term well-being and safety. [1]

MentalHealthBench is a new open benchmark for measuring how AI systems respond during realistic mental health conversations. [1] OpenAI co-created it with a global cohort of more than 80 licensed mental health experts from 22 countries. [1] The benchmark assesses safety, context seeking, preservation of user agency, and actionable guidance when appropriate, and its open release lets researchers examine the methods, conduct evaluations, and build on the work. [1]

Compared with earlier evaluations centered on emergencies and broad criteria, MentalHealthBench gives developers a more behavior-specific framework for examining responses across varied mental health situations. Its expert-designed criteria also shift attention from merely avoiding prohibited outputs toward seeking context, protecting agency, and offering appropriate guidance. The open release allows independent researchers to inspect and reuse the methods rather than relying only on OpenAI’s evaluation. What the sources don't address: how consistently the benchmark predicts safe, helpful performance across different cultures, languages, and deployed AI systems.