MAIN STORY
A more demanding test of difficult conversations
On September 23, 2026, OpenAI introduced MentalHealthBench, an open evaluation developed with more than 80 licensed mental health professionals from 22 countries. Rather than measuring only whether an AI system refuses dangerous requests, the framework asks whether it gathers context, distinguishes emergencies from ordinary distress, preserves autonomy and offers practical next steps. Its synthetic conversations range from family pressures and everyday anxiety to circumstances requiring immediate intervention.
The difference matters because emotional difficulties are not interchangeable. Someone grieving a setback may need a patient question rather than an emergency warning; a person signaling imminent danger needs a different response. Models should neither rush to diagnose nor obscure a crisis with reassurance. OpenAI says newer systems perform better on the expert-designed criteria. Those findings are informative but remain company-run evaluations, rather than proof of outcomes among users in their daily lives.














