For (1), I think it's handy to think of an encrypted zip file that contains a bunch of images. If you give it a valid password, it'll give you a corresponding image. If it produces an image that violates copyright, it clearly violates copyright, even if most of the images contained in it don't.
It's possible that if the model required sufficient instruction to reproduce Mickey Mouse ("create a line from (0,20) to (20,30) of thickness 2, then a line from ..."), and could not do so without that, then the argument wouldn't apply. And there's definitely a grey area once we're outside of the territory of reproducing images. But we definitely aren't anywhere close to that with these models, and looking at the "AI Safety" work, it doesn't seem possible to get these models to "behave".
> This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing.
Yeah, I agree that the plaintiffs aren't suing solely because the model can easily regenerate their images. But I think we'll find its insanely hard to get these models to avoid producing copyrighted images. They may be forced to license the training images.