It is absolutely not clear that an ML model is a derivative work. It might be for-profit but there's good arguments that it is incredibly transformative, and that each individual work the model is trained on is minimally important to the model (if you trained the model on every other document in the training set except the one being sued over, the model would perform very similarly). These are factors which will weigh against the copyright holder.