The most dangerous AI answer is not nonsense. It is the answer that sounds reasonable enough to end the conversation before you check it.