Relying on AI chatbots for health information is uniquely risky. We’re building an AI benchmark to make it safer.