That's true, but only in the sense that theft and copyright infringement are fundamentally different things.
Generating stuff from ML training datasets that included works without permissive licenses is copyright infringement though, just as much as simply copying and pasting parts of those works in would be. The legal definition of a derivative work doesn't care about the techological details.
(For me, the most important consequence of this sort of argument is that everything produced by Github Copilot must be GPL.)