If we can play semantics, the program (the compiled binary) can be used for anything with no field restrictions.
But the code is not the program itself, it's the recipe, and usage could be restricted in some specific ways.
In my opinion, since free licenses already have restrictions regarding distribution, saying AI models trained on this data are derivative works and must be licensed compatible (ie training data set, methods and models themselves being free).
I feel it's a better middle ground where the freedom of users are not violated nor restricted, and the code/knowledge stays free