If training LLMs on copyright protected art is itself Fair Use, then the motivating and only significant purpose of this is to adversely impact noninfringing uses. That’s the textbook reason for the Librarian of Congress to establish exemptions from the anti-circumvention rules, so even if as a legal hack that was effective initially, it wouldn’t be likely to last very long.

But I think whether this is a protection measure is undecided.

I think it would be hard to argue it isn't. It is the only value that such a system could theoretically provide, so there is no other reason someone would use it. What would be the argument that it is not one?

How are you applying a filter or circumventing anything if you are simply training a brand new model with a bunch of "cloaked" images?

Oh, i was thinking the scenario was more a specialized tool to remove the "glaze".

Hmm. That's a good point and now i am less sure.

I suspect it would come down to intentionality.

I looked at the definition in the law and it was unclear to me whether this technique met the definition. Among other things there's no "authorized deGlazing" to recover the original going on during human viewings. But there seemed to be enough leeway that maybe a judge could find a fit

Why would it need to specifically circumvent this instead of casually circumventing it simply by better mapping how the bits on a program map to how an image appears to the viewer?

if f(x)1011 and 1010 both "look" the same shouldn't a well designed AI learn equally well to recreate that particular look. It doesn't need a "deglaze" circumvention step it just needs a continually improving learning process that automatically accommodates both glazed and un-glazed images.

Right, a tool explicitly designed and used to "deglaze" images would be different from a network that just happens to not be affected by glaze, I expect.

Keyword being "effectively". If routine cropping and resizing defeat it, then it is hardly effective.