lilSorcerer said:
I would disagree here. Training the model entails feeding near "perfect information", a near "perfect copy" of a work to the model. This perfect copy is then in my opinion "abstractly stored" in the model (even though, of course the sequence of bytes representing the art is not literally directly in the memory of the model. Thus then means the owner has directly "used" a copy of a protected work. If they have rights, this is a non issue, but if not, it violates copyright.
Seems like a distinction without a difference . . . .
Human beings and an AI model, like Stable Diffusion, are trained on the same (or substantially the same) type of input data. For example, Stable Diffusion gets digital copies of all versions of all Mark Rothko paintings, ever. I have access to the same data, to the extent indexed on the web. And I also can see the real things in museums and private galleries.
In both of these cases, I have access to the "perfect copies" if not the real thing, directly or indirectly through the instrumentality of Stable Diffusion. And just as I might have access to a paint brush, paint, and a canvas (or many canvasses), so too do I have access to a computer with the horsepower to run stable diffusion. The only difference is that I have to paint with my hands in one case but I can prompt with my keyboard in another. The abstraction of how I map the patterns of "what is quintessentially Mark Rothko" may be different from how Stable Diffusion does it, and it almost certainly is. But the result is the same: Rothko is the inspiration for the future creation.
If the difference you want to draw is what happens to that input data once an AI model has been trained on it, that's not a compelling argument either. For example, if I paint something entirely of my own but in a style that is inspired by Rothko, would anyone dare challenge that my work product was my own? Absolutely not. Only a fool would, unless I created a forgery of Rothko's specific piece. In the same way, if Stable Diffusion creates art "in the style of Rothko" in response to
my prompt, the only difference is that I have used AI as the instrument through which my creative vision is executed.
In both of the scenarios described above, I have "used" a copy of "protected" work. The only difference is the instrumentality; in one case, the painting is by my own hand, but in the other the image I create is by a natural language prompt.
Although what is the difference between this process and our brains? Maybe there isn't much of a difference, and all human artists are regularly violating copyright as well? It's a strange but interesting thought indeed.
Either way, I think debating the pragmatics/dangers/benefits of ai is more important than debating the general morality.
The issue of whether using copyrighted images as training data for AI based text-to-image platforms violates copyright laws is different from what we were discussing above.