If the AI wants to be exempt from being punished for training on copyrighted works the bare minimum standard is that it doesn’t accidentally copy and reproduce large portions of that work. There are quite a few examples of Stable Diffusion doing this so I think even that low bar is met. I think anyone sane doesn’t want to create AI immunity to copyright damages even when the network is “accidentally” producing copyrighted output. If you can’t adequately control your technology to avoid that case you shouldn’t get immunity to liability for your problematic outputs. I get that testing for every possibility is impossible and that neural net explainability is far behind neural net technology so it’s very hard to proactively identify and debug these technologies. But just because those are hard problems doesn’t give you the right to steal from copyright holders.