OpenAI says deidentified data from the private sessions go into training. (Well, explicitly said they will not rule that out.) That changes a lot of the conversation.