Introducing MentalHealthBench

OpenAI has introduced MentalHealthBench, a new benchmark designed to measure how helpful and safe AI systems are during realistic mental health conversations. The tool was developed with input from experts and focuses on sensitive scenarios where a generic evaluation would miss important nuances. Instead of testing only factual knowledge or fluency, MentalHealthBench examines whether a model responds with appropriate support, avoids harmful advice, and recognizes when it should encourage a user to seek professional care. The benchmark evaluates both supportiveness and risk. That dual focus matters because a response can sound empathetic while still giving dangerous guidance, or it can be cautious but fail to offer useful help. MentalHealthBench provides a structured way to test models across difficult conversations, including moments of distress, confusion or vulnerability. It is intended to help developers identify weaknesses before deployment and to track progress over time. The release reflects growing concern about how people use AI for mental health support. Chatbots are accessible at any hour and can feel private, which makes them attractive to users who may not have immediate access to a therapist or crisis service. But they are not a replacement for trained professionals, and they can cause harm if they mishandle serious situations. Benchmarks like MentalHealthBench aim to make those limitations more measurable. They also raise broader questions about responsibility. If a model is deployed in a health-related context, what level of safety should be required? How should companies handle crisis situations, privacy and escalation to human care? MentalHealthBench does not answer every policy question, but it gives researchers and regulators a common reference point. As AI becomes more integrated into daily life, evaluations that combine empathy, accuracy and risk awareness will be crucial for building systems that support people without misleading or endangering them.