But I'm also a musician/artist and so I find some of these conversations odd. The problem with them I see is that they are oversimplified. To get better at drawing I often copy other works. Or I'll play a piece exactly as intended. Then I get more advanced and learn a style of someone I admire and appreciate. Then after that comes my own flair.
So I ask, what is different between me doing it and a machine? The specific images being shown in this article shows that Hollie is doing the same process as me and the machine. Their work is in fact derivative. There's absolutely nothing wrong with that either. They say a good artist copies but a great artist steals. I don't think these generators are great artists, but they sure are good ones.
I learn by looking at a lot of art and styles. I learn by copying. I am able to be more efficient than the machine because I can understand nuance and theory, but the machine is able to do the former much better than I can. It can practice its drawings a hundred or thousand times an hour.
Now there are highly unethical stuff that is actually going on and I don't want the conversation getting distracted from. Just today a twitter account posted a demo of their work and actively demonstrated that one can remove watermarks with their tool[0]. This is bold and borderline illegal (promoting theft). There are also people presenting AI generated work as human digital paintings (we need to be honest about the tools we use). People presenting work in ways that it was not actually created is unethical. But there are other generative ethical concerns.
Now there are concerns about photographer's/artist's rights. If I take someone else's work and post it as my own, that is straight up theft. Even celebrities can't post photos of themselves that were taken by others[1]. This gets muddled if I make minor changes but it's been held up in court that the intention matters and it needs to be clear that the changes were new in an artistic manner and not a means of circumventing ownership rights. These are some of the bigger issues we're running into.
A problem with these generative models is interpretation. How do you know if the image you produced actually exists in the wild or if it is new and unique? There's been papers that show that there are privacy concerns[2] and that you can pull ground truth images out of the generator. I'd argue that this question is easier to answer the more explicit the density your model calculates. Meaning that this is very hard for implicit density models (such as GANs), moderately difficult for approximate density models (such as diffusion and VAEs), but not too bad when pulling from explicit density models (such as Autoregressive or Flow based models).
This is a concern that is implicit by articles such as this, but fail to actually quantify the problem here: "How do we meaningfully differentiate generated images from those made by real people?" I'm a strong advocate for the study and research for explicit density models, but a good chunk of the community is against be (they aren't currently anywhere near as powerful, but there's nothing theoretically in the way. I'd argue it is that few people are researching and understanding these models. There is a higher barrier to entry). So I'd argue that the training methods aren't the major concern, but what is actually produced. While the generators learn in a similar fashion to me it is clear that I'd get in trouble if I was passing off a fake Picasso as a legitimate one. But it is also fine for me to paint something in that same style as long as I'm honest about it.
The nuance here really matters and I think we need to not lose sight of that. This is a complex topic and I would like to hear other views. But I'm not interested in mic drops or unmovable positions. I don't think anyone has the right answer here and to solve it we must get a lot of different view points. So do you agree or disagree with me? I especially want to hear from the latter.
[0] https://twitter.com/ai_fast_track/status/1587475575479959559
[1] https://collenip.com/taylor-swift-entitled-say-photographers...