> They merely blend multiple copyrighted works together to create a new derived work Is this true though? If your model is trained on multiple copyrighted wor

> They merely blend multiple copyrighted works together to create a new derived work Is this true though? If your model is trained on multiple copyrighted works to learn the statistical probabilities and relationships between tokens of those works, because they represent examples of common language usage, that's not storing a copyrighted work en masse. If those probabilities happen to pump out a chunk of copyrighted work it's due to the statistical relationship between the tokens (overfitting?), not that it's saved that specific content.