Comment by manholio

But they are not original works, they are wholly derived works of the training data set. Take that data set away and the algorithm is unable to produce a single original pixel.

The fact that the derivation involves millions of works as opposed to a single one is immaterial for the copyright issue.