If models can learn for free, then the models (training code, inference code, training data, weights) should also be free. No copyright for anybody.

And if you sell the outputs of your model that you trained on free content, you shouldn't be able to hide behind trade secret.