LLMs are not optimized only for truth they are optimized for a more complicated objective that includes e.g. humans liking their output. It is a universal truth that to get humans to like you, you have to lie to them.