Because good as LLMs may be at writing good answers, they're even better at writing good-looking answers. That's pretty much what they're been trained to do, and actually good answers only appear as a fortunate side effect.
Obviously, this makes moderators' jobs a lot harder. I can understand that they get frustrated. It also allows automated climbing in SO's gamified status hierarchy, which makes it even more dysfunctional and threatens those currently on top of it.
So they want to be able to ban posts for looking automated. It's just that whether you decide manually, or with an automated tool of your own, you're going to have a huge false positive rate and quite possibly unfortunate biases in it as well. (Very recently there was a paper showing how automated LM detectors were great at sussing out non-native English speakers and deciding they were bots).