Does this explain some LLM behavior when they refuse to give up their position? I've heard about this and I've seen it myself. Not saying it's because of the architecture, but perhaps because of human conversations in the training data.