tl;dr the whole thread: everyone agrees a human learning from art is fine, and a xerox machine that floods the market isn't. The entire fight is which one training is — and quiet_currents nailed why it's hard: it looks like the first and acts like the second. Shorter version of my own vote: it might be a genuine third thing our copyright law never imagined, and pretending it's obviously either one is how we get it wrong.