Such AI is still limited by the same resolution (480 interlaced) and it can't create detail where it finds none.
Nope. The whole point of these systems are to fill in missing data points in images. Subpixel data can be recovered temporally (reviewing several adjacent frames), predictively (half a face gives enough details to figure out the rest because faces ain't that different) or reconstructively (building a catalog of people, props and sets and figuring out when each person or item is in a frame and using close-ups from an entirely different episode to fill in detail).
All of these methods are in use now. They are not likely to be ready for something like a DS9 remaster...yet. Someone out there is probably writing a PhD thesis on how to do it as I type here.
"AI" gets a lot of hype these days, much of it overblown, but the tech can do image processing. Image processing is precisely how self-driving cars operate.