Collage may be a better analogy than paint. An artist can make an image with a collage of bits of other artists' works, in much the same way that a generative network can make an image out of image elements derived from other artists' works.

The question is not whether the "AI" has any ownership, but whether the original artists do, or whether the programmers do, or whether the people selecting training images do, or whether the person selecting a final artwork out of a mass of inferior generated images does.

> An artist can make an image with a collage of bits of other artists' works, in much the same way that a generative network can make an image out of image elements derived from other artists' works.

That's not a good analogy because a human is involved. The entire point here is whether living entities other than persons can have copyright, and the answer (in the US) is no.

"living entities other than persons" having or not having copyright is unrelated, as programs are not living entities. Humans are definitely involved in developing, training, operating, and choosing output from machine learning art-generating systems.

The AI didn't write itself. A human chose the algorithms, the training data, the goals, etc....

Apparently that's nothing. As if all this AI generative art is trivial.

What you say makes sense in principle, but I don't think it sufficiently accounts for the fact that with AI we can mass generate enormous quantity of artwork with essentially zero incremental effort.

If I create an AI that creates billions of different image of dogs sitting on a park bench, and then someone else creates an image of a dog sitting on a park bench, odds are pretty high that it would look substantially similar to at least one of my AI generated images. Would it then be fair for me sue whoever is using that image for copyright infringement?

You certainly can sue, but you would be expected to fail because accidentally looking substantially similar is not copyright infringement.

There is some precedent for substantial similarity in certain popular songs, but it is based on convincing the court of the argument that the author of the later work must have had heard the original one (for example, over the radio) and copied parts of it, possibly unintentionally.

However, if it's reasonable to demonstrate that in your proposed case that "someone else" had not seen that particular one-in-a-billion image of dogs sitting on a park bench (since probably noone had ever looked at that particular image out of the billions you generated), then their creation is not a copy of your image ("bits have color" - even if it would happen to be perfectly identical, it's not a copy if it was not actually copied) nor derived work of your image and so it does not infringe upon your copyrights no matter how similar it looks.