Even in your latter paragraph, it wouldn’t be an infringement. Assuming the art was lawfully accessed in the first place, like by clicking a link to a publicly shared portfolio, no copy is being encoded into the model. There is currently no intellectual property right invoked merely by training a model-- if people want there to be, and it isn’t an unreasonable thing to want (though I don’t agree it’s good policy), then a new type of intellectual property right will need to be created.
What’s actually baffling to me is that these pieces presumably are all effectively public domain as they’re authored by AI. And they’re clearly digital in nature, so wtf are people actually buying?
No, this isn’t really correct. The US Copyright Office has released policy that pretty clearly states where the line falls and it’s certainly beyond super simple prompts. In fact, by the reasoning in the policy document, I’d say it’s any time where if the AI were replaced with a human and you’d want a work for hire agreement to assign copyright, then that is likely non-copyrightable subject matter.
I’ll add, how this works with modern AI art flows, still remains to be seen, but I think probably on the side of no copyright. Currently, works use very elaborate prompts, some edits, bashes, and masks in an editor and then img2img and inpainting to really get your work where it needs to be. However, under the current rubric, the sort of nexus of creativity is still happening in the model so unlikely to be granted copyright.
The Copyright Office recently released a webinar on just this point. Basically anything that is creative and human generated is still granted copyright, but the AI generated components are themselves non-copyrightable. In your examples, those components are fairly de minimis (small and insubstantial) and so the overall copyright of the work wouldn’t be impacted.
It could of course go up to the scotus and effectively a new right be legislated from the bench, but it is unlikely and the nature of these models in combination with what is considered a copy under the rubric copyright in the US has operated effectively forever means that merely training and deploying a model is almost certainly not copyright infringement. This is pretty common consensus among IP attorneys.
That said, a lot of other very obvious infringement in coming out in discovery in many of these cases. Like torrenting all the training data. THAT is absolutely an infringement but is effectively unrelated to the question of whether lawfully accessed content being used as training data retroactively makes its access unlawful (it really almost certainly doesn’t).