I wonder if future models can recognize which are the type of information that is better censored in host vs in training, and automatically adjusts its model accordingly to better fit with different user's needs.