I don't think this is as cut-and-dry and you make it out here. If I train a model on, say, every one of New York Times' and release it for free and it finds use as a way of circumventing their paywall I have difficulty justifying that as fair use/fair dealing. The purpose/character of the model should indeed be a factor but certainly not nearly as dispositive a one as I think you're suggesting.
To the extent that training the model serves a research purpose I think that the general use / public release of the trained model does not in general serve the same research purpose and ought to have substantially lower protection on the basis of, e.g., the effect on the original work(s) in the market.