Is an ML model binary file created by using copyrighted work as its learning data, a derivative work of the copyrighted work? I don't think so.

If one takes another's work, cuts it up and makes collages out of it, however multidimensional, what is the piece size threshold that makes the collage non-derivative?

This is the most fundamentally important question of AI-related law, and nobody knows the answer as it hasn't been tested by any court AFAIK (at least not in the US).

We already know the tribunal will take AI's side, not because of any justice or ethical reason but because of capitalism. They will interpret the law as saying whatever they want the law to say.