Overall, OpenAI found that "mental health conversations that trigger safety concerns, like psychosis, mania, or suicidal ...