I tend to lean towards the idea that when you train an ML model on a very wide range of content that the resulting model weights should probably fall under free use (or local equivalents), but that the outputs of said model do not necessarily do so.