> In copyright law, a derivative work is an expressive creation that includes major copyrightable elements of an original
The key part here is "major copyrightable".
To be clear, my argument doesn't show that statistical generative approaches to art are violating copyright by using copyrighted data. It only shows that the argument "humans learn on copyrighted stuff therefore it's legal for AI to do it" isn't a sound argument.
Another argument that relies on the definition of derivative work might be able to show that you can train a model on copyrighted data without producing works that violate copyright.
"humans learn on copyrighted stuff therefore it's legal for AI to do it" - Learning can never be a violation of copyright. A work (digital or physical) can potentially violate copyright.
So there are two potential "work" here:
1. The model itself
2. Images created with the model.
I would argue that (1) can't violate copyright as it's not a derivative work
(2) may or may not violate copyright but if it does, then it does so to exactly the same degree as it would if it wasn't AI generated.
In other words "You're not allowed to plagarize copyrighted works no matter what production method you use".
We can break this into (1) and (2):
(1)
The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse.
To preempt a few (obvious) arguments about what it means to be "transformed", keep in mind an encrypted zip file containing images of Mickey Mouse would not be considered transformative, even if it happens to contain images that aren't copyrighted.
(2)
> then it does so to exactly the same degree as it would if it wasn't AI generated.
This is the conclusion of an argument, but I'd need to see how you show that "it does so to exactly the same degree". Right now, it looks like you're taking that as a base assumption.
This is a strong argument and gave me pause. I think if Disney sued they could probably win.
However - examples such as this are probably quite rare. Mickey Mouse is so uniquitous that is massively overrepresented in the training and data and that has resulted in overfitting. For the vast majority of artists and almost certainly for the plaintiffs in this case, the demonstration wouldn't work. You most likely cannot produce an infringing image from a concise text prompt.
Arguably one of the outcomes of legal action might be the need for a lot of effort to remove examples similar to the Mickey Mouse example. This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing.
> (2) but I'd need to see how you show that "it does so to exactly the same degree".
Not sure I follow? If I produce an infringing image and a court agrees it's infringing then I'm liable. It doesn't matter how the image was produced. I don't see what there is to dispute here?
For (1), I think it's handy to think of an encrypted zip file that contains a bunch of images. If you give it a valid password, it'll give you a corresponding image. If it produces an image that violates copyright, it clearly violates copyright, even if most of the images contained in it don't.
It's possible that if the model required sufficient instruction to reproduce Mickey Mouse ("create a line from (0,20) to (20,30) of thickness 2, then a line from ..."), and could not do so without that, then the argument wouldn't apply. And there's definitely a grey area once we're outside of the territory of reproducing images. But we definitely aren't anywhere close to that with these models, and looking at the "AI Safety" work, it doesn't seem possible to get these models to "behave".
> This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing.
Yeah, I agree that the plaintiffs aren't suing solely because the model can easily regenerate their images. But I think we'll find its insanely hard to get these models to avoid producing copyrighted images. They may be forced to license the training images.
You, and only you, are liable.
If someone was distributing a password protected zip-file where if I type in "Mickey Mouse", I get a photo of Mickey Mouse, the zip file violates copyright. It doesn't matter that the underlying representation of Mickey Mouse is distorted in storage and recovered later. It doesn't matter that in order to retrieve the image, I have to go through a series of operations.
Of course, if I opened the zip file and put the image on a t-shirt, I would also be violating copyright. But the zip file itself is copyright infringing.
At this point we can just wait until the courts clarify this for you because you seem to be missing a fundamental aspect of how our legal system operates. I’d much rather go and watch football.