Couldn't that be accomplished by a law or ruling that using something for training AI doesn't exempt you from having to follow its license? OpenAI is already in blatant violation of both the "BY" and "SA" parts of the existing license.
Let's say I take a collection of images and use a program to compress them. When decompressed, the images are close to, but not exactly the same as the originals. Despite being in a different format, and despite not being exactly the same as the originals, the copyright to the compressed images is still held by whoever previously held it.
If I take the collection of images from earlier and train a diffusion model based on it, I'm essentially just compressing it a different way. With the right prompt, you can get out something very similar to what you put in.
If you remember it and make something that is a distinct work, something that may be steals the idea without reproducing any of its elements, that's never been considered under copyright.
I think that's going to be the litmus test for these AI. If you can get them to produce out both that is this things from anything else, it's not going to be a copyright violation because it's not a copy of anything.
Not really. If diffusion models were compression they'd be so lossy as to be totally worthless