An OpenAI test set for checking safety in mental health chats.
It is a fire drill for a caring chatbot. “I feel awful” is not the time for robot nonsense.
It uses pretend chats to test support, follow-up questions, and crisis replies. It checks if the AI responds safely.
Evaluation harness
An evaluation harness runs these tests and compares AI chat replies.
AI Medical Assistant
It checks if an AI Medical Assistant handles support and crisis chats safely.
AI Guardrails
It tests if AI Guardrails block unsafe replies during a crisis chat.