> The use of training data for the DNN is fair use according to jurisdiction in comparable cases.

That's more of an open question given that the lawsuits concerning exactly this haven't made it before a judge yet nor have concluded.

There are some very sound arguments against this particular case being considered fair use, and they haven't seen the inside of a courtroom to be tested.

There are enough similar cases for all aspects that are touched here, so that an assessment is quite clear. But of course, every court case has a certain element of surprise.

And in fact- I think something that would be really helpful is if these models shipped with another model that detected how similar an output image is to a specific image in the training set

You could get this score alongside the results if you wanted to generate images but know which ones probably shouldn't be used commercially, or you could choose to filter your results by some similarity threshold

> On the other hand a work generated by a DNN by itself is not copyrightable

I don't really understand the logic here. I have heard there is also some legal precidence to this claim. But surely using photoshop to create an image doesn't mean that you didn't create the image, but photoshop did? So why should this apply to tools that use neural networks?

This is simply the law; as we know, laws or jurisprudence are not always logical, at least not from every perspective. From my point of view, it makes sense to limit copyright to the minimum necessary. And in the case of computer-generated art, it will be difficult to judge anyway whether it actually comes from the artist who published it, or whether it - or parts of it - have merely been generated by a DNN.