I did some ratings on chatbot arena and i noticed one things. When an ai honestly said “i dont know that” or “i dont understand that” it was always better received by me and felt kinda smarter.

Does some dataset or lora train on that? Or is “knowing about not knowing” too hard to achieve?