OpenAI has introduced MentalHealthBench, a new benchmark aimed at evaluating AI responses in the sensitive domain of mental health.
According to the announcement, the benchmark is expert-informed and focuses on two core dimensions: whether AI responses are helpful and whether they are safe.
The evaluation is conducted across realistic mental health conversations, reflecting scenarios that users may actually bring to an AI system.
The company framed the release as a step toward more rigorous assessment of AI behavior in high-stakes contexts, though no specific scores, model names, or technical details were included in the source text.