The dislike button is if nothing else a sanity check for someone to have control if your model goes haywire. Realistically it will go haywire in a non-trivial number of cases. If every month your model goes haywire for 1% of users then you may see dislike used 1% of the time. That's tiny, right? However without it those users may leave and then next month another 1% will be hit and so on. Users remember the bad things a lot more than the good things so they will remember your AI f-ups and the lack of a way to fix it.

The problem with data driven UX is that the data is often short term (ab tests are expensive) and doesn't cover long term complex user impact.

Edit: For example, I’ve seen a lot of Etsy users complain about their front page being filled with NSFW recommendations and no way to get rid of them. I doubt they’ll be using Etsy much in the future.