Leaked Meta AI Guidelines Show How the Chatbot Is Trained to Spot Sensitive Content and Avoid Controversy