I'm guessing under the hood these LLMs are told to agree with a user if they tell it some response is incorrect. When I'm generating template code I find this happens a lot. From the ground up designed to fool a human in a very convincing way.