Most images found on Wikimedia sites actually require attribution to the original author, including for derivative works. So if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output. (SVM is one machine learning method that can easily do this, but I'm not sure if it's ever been applied to image generation.)

> if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output.

This is a huge leap that would change copyright law dramatically, even for humans.

Youre asserting that copyright doesn’t just protect from reproduction, but from use as an example to learn from.

I think this would break copyright altogether. What musician or artist can say a work is 100% original and does not incorporate any details or concepts from work they’ve seen / listened to / read?

I know people hold that position, but it’s important to be transparent that applying copyright to training/learning (in addition to output) would be an overhaul of copyrigh, not a clear and obvious application of the law as it stands.

You only have to make that leap if you suppose that machine learning algorithms actually learn and then become capable of creating their own original works, rather than producing mechanical transformations of their inputs.

Which, I don’t think that’s an inherently absurd statement. I mean, at some point humans physically exist, so a sufficiently complex artificial being could exactly replicate our physically processes, and it would be hard (imo) to argue that such a thing isn’t a person.

But I don’t see the non-big-leap option. Either we admit that machines can actually perform creative works which are their own ideas, or we call the thing they do a derivative work, and we admit that we can’t just eyeball the difference between an original work and a derivative work anymore. That there’s something intrinsically special about having a human in the loop.

> You only have to make that leap if you suppose that machine learning algorithms actually learn and then become capable of creating their own original works, rather than producing mechanical transformations of their inputs.

Why do you think humans can do that, and that what we call “original works” of humans are not just mechanical transformations of the human’s inputs.

How, in a universe of deterministic physical laws, is it possible for a human to do anything else?

That seems pretty close to what I said in the next paragraph, just approaching from the other direction.

> including for derivative works

This presumes that an input having influence on an output means that the output qualifies as a derivative work of that input. A work being in some way derived from another work does not mean that it is a "derivative work" for the purposes of copyright law. Many artists want to make this presumption, but IMO it's extremely dubious as a general claim. It is likely that when this all shakes out, specific works will be able to be established to be derivative works of specific other works, but simply being an input will not cause that. (caveat: IANAL, but I do have a law degree and IP experience)

> So if you wished to include such images in training, you would also need an explainable AI that could tell you exactly what training inputs are influencing its output.

That would be like requiring human artists to enumerate exactly every single influence on each piece of art they make. Did they see a painting by so and so at six years old that had a 0.1% influence on them? Better include it too.

Publicly-licensed works are 100% copyrighted. That's why they need licenses. A true public work doesn't need or get any licensing. Linux is copyrighted, and licensed, to become f/oss. Shakespeare is actually public domain, no license required.