But does that still hold when the model memorized a chunk of the training data? Or can a network plagiarize output while being a transformative work itself?