In the baseline model we use unsupervised learning or just the Internet content, but for the chatgpt side, we introduce the human labeler. How do we make sure this part is part of the known?