Once you start recreating copyrighted works from hashes this analogy becomes relevant, until then how can you compare the two when the distinguishing feature is

Once you start recreating copyrighted works from hashes this analogy becomes relevant, until then how can you compare the two when the distinguishing feature is its ability to reproduce the training data.