ML works well, when it is only trained on correct data.

ML trained on all the things ever written about programming is going to get things wrong quite a lot of the time, and you'll only have the context of its reply to you, for you to figure out whether the ML answer is correct. And of course, ML always wants to please you with an answer.

SO, on the other hand has accepted answers, and upvoted answers that are easier to decide on whether they're correct (or what you're actually looking for). And there's more nuance to it as well.