If the training toggle is switched on, maybe OpenAI doesn't consider a chat to be private? Therefore making this a 'safe' statement.