> "It is certainly not clear in a moral or legal sense whether training constitutes making an unauthorized copy." Another complicating factor here is to what e

> "It is certainly not clear in a moral or legal sense whether training constitutes making an unauthorized copy." Another complicating factor here is to what extent the mainstream definition of "training" even applies to things like LLMs. The mismatch between what "training" typically means vs. what it means very specifically in a ML context is the source of many problems in this discussion. We have for example heard many arguments that a human listening to every Beatles album, and then writing a new song based heavily in the style of the Beatles, is very clearly legal. But it's not clear that the ML model "learns" or "trains" in any way that is analogous to the human mind. Another problem that I think we're not really contending with is to what extent LLMs generalize and to what extent they memorize. The more memorization that occurs the more compelling the argument will be that this is to some extent copyright-violating. It certainly doesn't help the generalization argument that many LLMs can be compelled fairly easily to reproduce parts of their training data verbatim.