> "computer generated art" is also not protected under copyright
These AI's mix art that could be under copyright and very different licenses (I've seen watermarks while using them). You can assign any license you want to the output, but that doesn't mean the owner of the original art won't find a similarity and sue you. So basically it's a risk. Laws are blind, there is no automatic winner in a legal dispute, what you can prove in court and the laws that apply to your jurisdiction are what count, and also how much money you will spend in the legal process whether you are right or wrong.
Not a lawyer! (tm)
As a tangential topic, all those AI companies that profusely talk of using AI ethically must first disclose their data provenance. Otherwise, it is just lip service.
Having said that, I could hypothetically use an open source compiler/linker to compile copyrighted source code and produce executable artefacts. That doesn't protect me against infringement complaints.
In this hypothetical case, the compiler/linker creators are not responsible for the infringement; As the user, I would be.
The situation is murkier with AI models. The model itself may have been created from sources of unknown provenance (i.e., including potentially copyright infringing material). Most AI models are released with licensing that puts the onus of the legality of the generated output on the user. This is slightly different from the open source compiler/linker example above because the user is ostensibly in the dark about the source of infringement.
There is also the question of concepts of fair use, transformative works and derivative works that is in the legal domain. All we can say at the moment is that this is untested in courts and there will be some precedent-setting caselaw that will emerge from lawsuits that will make it to the courts in various jurisdictions around the world.
So, tread with caution. Use models that are open about their training data and algorithm provenance.