> Is that settled law?

It seems to be the fairly consistent approach of trial courts addressing the question under different soecific fact patterns in different contexts; its not “settled law” in the sense of nationally binding precedent (which would take either a Supreme Court ruling kr separate appellate rulings in every circuit).

> And IMHO, AI training violates the spirit of fair use. It's not really a method of criticism or commentary.

Plenty of transformative uses that have been held to be fair use are not criticism or commentary, and AI training is a transformative use where, for any individual work used, the end product is both a very different class of work and the used work indiviudally has very small impact on the final work.

> It's a system to use people's own work to undermine their ability to economically subsist on that work.

Courts seem to disagree that this is generally the case with AI training as such. (And AI training being fair use would not make the use of models to create works that would otherwise be infringing copies with that function through inference any less infringing.)

> Though how about this for a proposed exception: you can AI train on anything as fair use: only if release your model and weights public domain open source.

You are, of course, free to try to convince Congress to amend copyright law to apply that rule (though since the current statutory form of the fair use rule is itself a legislative adoption that follows pre-existing court rulings on fair use as a Constitutional limit on the copyright power stemming from the First Amendment, Congress may not actually have the power to narrow it that way.)