I draw the line at edits that consider semiotic meaning. Edits are acceptable if they apply globally (e.g. color correction to compensate for faded negatives), or if they apply locally based on purely geometric considerations (e.g. sharpening based on edge detection), but not if they try to decide what some aspect of the image signifies (e.g. red eye removal, which requires guessing which pixels are supposed to represent an eye). AI makes no distinction between geometric and semiotic meaning, so AI edits are never acceptable.
Easy counterexample: dumb unsharp masking will ruin close-up scenes that are shot for softness and/or have bokeh. ML upscalers can do this too when applied mindlessly. But you can also train an upscaler on the same type of footage, or even on the parts of the same footage available in higher definition. Even if you don't, matching the upscaler with the intent behind the content is your job.
The separation you're talking about is imaginary, the line doesn't exist. Any tool will affect the original meaning if it doesn't match the execution. Remastering is an art regardless of the tool, and it's always an interpretation of the original work. It's fine to like or dislike this interpretation.
Remastering can screw up intent with something as simple as color grading.
But there is a line here. An editor that's using simple tools knows exactly what they're changing, and if they're using simple frame-global tools then they're not introducing anything that wasn't already there.
If you throw an AI at things, it will try to guess what things in the image are, and make detail adjustments based on that.
So that's three categories of edit, easily distinguished: human making frame-global changes, human deliberately changing/adding details, AI changing/adding details in a way that's basically impossible to fully supervise.
It sounds like they accept category 1 in remastering, even though it's not foolproof, and reject 2 and 3.
No, "AI" is absolutely not uncontrollable magic that does something you don't want sometimes. It's not an issue really, you always have arbitrarily granular control of the end result, with ML tools or not. You can train them properly, you can control the application, you can fix the result, you can do anything with it. It's the usual VFX process, and it's not the only tool at your disposal.
The problem is that remasters don't make a lot of money, so instead of a properly controlled faithful representation (or a good rethinking) it's typically a half-assed job with a couple filters run over the entire piece. Another issue is that you now have two possibly conflicting intents - one from the original and another from the remaster. ML haven't changed anything in here, it's always been like that.
Sure, my point is that proper remastering is not just applying a couple ML filters. If you're doing that you should either do this selectively or fix the result by other means, i.e. the same thing you would do with dumb processing. This is a labor intensive VFX work, feasible for a new movie but not feasible for a remaster.
Yes, back in the mid-late 80s Turner Entertainment colorized a huge number of old films in their vaults to show on cable movie channels. It was almost universally panned. It was seen at first as a way to give mediocre old films with known stars a brief revival, but then Turner started to colorize classic, multi-award-winning films like The Asphalt Jungle and the whole idea was dismissed as a meretricious money-grab.