That's really no reason LLMs can't be all that though, and I don't mean instructed to. The sycophantic ass-kissing is a consequence the post-training reinforcement learning to be a 'helpful AI assistant'. Base models aren't really like that at all. Hell the original Bing ignored a lot users and would often refuse to entertain further requests if you 'upset' it. Microsoft wasn't telling Bing to do all that. In fact, they replaced the model for it.