This ruling has next to nothing to do with generative AI.

If anything, the fact that there was enough ambiguity that it took 35+ years for a court to decide that someone tracing a photograph was, in fact, not fair use is a good indication that it will be very hard to use current laws to stop outputs of generative AI that have been trained on copyrighted material.

I think parts of the ruling has a lot to do with generative AI.

The big problem for Warhol is that people paid for the image he produced instead of paying for the original photograph for use as a reference to Prince's likeness. So the works were for the same purpose.

If you use generative AI to produce a 3d State of Liberty for your video game I doubt you'll end up losing a case by a photographer of the Statue of Liberty because the purposes are actually different. But if you are using generative AI of stock imagery to produce stock imagery I think you're going to run into the same problem as Andy Warhol's foundation did.

Based off of current copyright statutes, nothing about generative AI creating stock photos by being trained on stock photos would violate copyright law; only if it was substantially similar to an existing work and was not transformative of that existing work. Just because the AI (read: the human) has seen work doesn't mean that creating a photo with the same subject and even with the same framing and techniques is inherently infringement.

> only if it was substantially similar to an existing work and was not transformative of that existing work

That's a giant only.

If the resulting image is nothing close to the source image then yes I agree you're unlikely to lose any copyright case.

But its the cases where they are similar that anybody cares. I pick the word "stock imagery" for a specific reason; I do no see a way that you can have a transformative use of stock imagery since its use is so vague. And a reminder (with a quote this time) transformative does not refer to the production it refers to the use.

If you painted every single stock imagery you could find and let companies license those paintings (or pictures of your paintings) instead of using the source stock imagery I believe you'd lose a copyright lawsuit.

> [1] AWF contends that the Prince Series works are “transformative,” and that the first fair use factor thus weighs in AWF’s favor, because the works convey a different meaning or message than the photograph. But the first fair use factor instead focuses on whether an allegedly infringing use has a further purpose or different character

[1]: https://www.supremecourt.gov/opinions/22pdf/21-869_87ad.pdf

Training a model on a bunch of images by {artist}, and then telling it "Generate me 20 images in the style of {artist}" seems pretty close to this ruling.

It's non-transformative in the legal sense that {an original image by X} and {a generated image in the style of X} are likely used for the exact same purpose (be it display, advertisement, whatever).

I think it would need to be a bit more than just that.

If you have a Picaso style advertisement for Ford its probably transformative w.r.t. any of Picaso's works since none of them as-is could be used for that advertisement. There's no way Picaso painted any Ford motor vehicles.

As well as if you used those imagery as an example of the generative capabilities the originals wouldn't suit that purpose. Or if you had a collection of AI generated art then the originals also wouldn't work.

> Training a model on a bunch of images by {artist}, and then telling it "Generate me 20 images in the style of {artist}" seems pretty close to this ruling.

No, it's not close at all. Did you look at the images in question for the ruling? Warhol's picture is very clearly a tracing of the photograph. Unless you're using the AI to generate an almost-exact copy of one of the training images, this ruling does not apply.

The decision is not an infringement decision. AWF did not dispute the finding that the images are substantially similar.

> The decision is not an infringement decision

> Supreme Court rules Andy Warhol’s Prince art is copyright infringement

And fair use is the main exemption to copyright, and that was decided here.

No it wasn't decided there. There is no analysis as to whether the image is infringing or not and it's not even a debate, because the warhol foundation admitted the pictures were substantially similar.

Even better for AI, then, if the supreme court was divided on whether such a small change is a valid use even with similarity assumed.

> Training a model on a bunch of images by {artist}, and then telling it "Generate me 20 images in the style of {artist}" seems pretty close to this ruling.

Maybe, but that is not how dalle and other generative AI work right now it is trained on a huge number artists and images. Not just mixing and matching from images from one artist even if you ask for something in one artists style. Generated an images that falls under copyright seems possible, but automatic.

I was actually watching this case precisely because it has implications for generative AI

That's a complete misunderstanding of the issues present in this case, which were very nuanced. I say this as a copyright litigator. And I'm happy to explain further if you'd like.

Please share your thoughts

Except that generative AI is actually generative. It isn’t recreating a work by shape and line and form. It is using the knowledge gained from studying many works to determine when given a random pixel, what color pixel should it and its neighbors transform into to best fit the requested parameters.

Stable Diffusion starts with random noise and attempts to denoise that random information into a coherent picture. That it has in the past learned that certain types of pictures should have say something resembling a watermark is just it being incredibly dumb but ultimately no different than it learning that horses should have something resembling manes or people’s hands having something resembling fingers.

You are just wordsmithing around the value of, in this case, the orange stencil-work, when you say "AI is actually generative"

Generative AI isn’t doing anything at all like what Warhol was doing here, and there is nothing new in the terms of broad rule in the decision as to what is “transformative” that would signal much as to how the arguments relative to training and use of generative AI might play out.

Category error: generative AI isn't making a copy. Says so in the name.

The threshold in law is "substantial similarity", not an exact copy.

"In the style of..." is not substantially similar enough IMO.

I think a better foundation for generative AI regulation is the moral rights of the artist [1], not copyright law.

But the courts will take years, if ever, to get that far.

1. https://en.wikipedia.org/wiki/Moral_rights#In_the_United_Sta...

Geez - we went from being a category error, to something you disagree with based upon your opinion... what next?

You still have a different type of category error: The creator, AI in this case, lacks legal personhood, hence cannot be considered the legal author nor hold copyright.

That's not a category error, its a trivial characterization based upon the marketing term used to advertise the kind of service and is not a legal argument in the slightest.

Training the model can still be argued fair use. Using the model to recreate copyright works is not.

How does this ruling impact generative AI?

(IANAL) The argument is probably something along the lines of this ruling raising the bar on making a fair use argument for transformative use (which would presumably be one of the defenses for using copyrighted data for training). In practice, I have to believe this is a sufficiently specific case and has sufficiently closely related precedence (like the Obama Hope poster) that I'm not sure this says much with respect to generative AI.

It makes it harder to prove that what the AI generated is transformative work.

How so? This just looks similar to the Obama photo situation. You look at the original and then you look at the artist rendition and in both cases they line up pretty closely to the original. The transformative work being done by the AI is way more than this unless the artist intentionally is trying to preserve the original closely. In both these cases, the artist is making decisions to not be that transformative from the original.

Also, I can do this all day long if I am not selling my output.

I haven't really been following the AI stuff that closely, but in a similar conversation last week someone was noting that while a lot of the model creators said the original images aren't there, in some/many cases they actually are and if you ask correctly the exact (or close to it) image comes out the other end.

I don't know the details, or enough to have a real opinion on it, but it seems like if you have a system that you can request something from and an input image comes out mostly intact on the other end, that shares a lot of similarities to a database and lossy compression. If that's actually possibly, then I have questions about how much of an original is used in the output even if it's changed, and at what point it becomes transformative.

Again, I'm not sure about most of this, and don't even have the source comment I'm remember, much less their source for their assertions, but I do have lots of questions and suspicions, if indeed what we think we know about these models based on prior statements ends up being wrong in some cases as we learn more.

transformative doesn't refer to modifications done to the work of art.

Transformative refers to how the works of art are used. For example, if I make a compilation of every single Barry Bonds homeruns the clips of the games are copyrighted but the video I produced isn't a substitute for an actual baseball game so the resulting work is transformative even if none of the video footage has been altered.

Maybe. I wouldn't bet on it if the copyright holders of the clips objected.

I think legally you'd be fine [1].

But certainly the copyright holders would succeed in getting your video taken down from YouTube since it wouldn't be a DMCA request. (Fair Use doesn't mean a private company must host your video).

[1]: https://www.citizen.org/article/a-guide-to-fair-use-in-posti...

I don't see how that definition and this court case make it harder to prove that AI generations are not transformative work.

The point I'm trying to make is a point of clarification as you seem to have a misunderstanding of what transformative means w.r.t. Fair Use since your post [1] talks about comparing the differences between the source work and the AI work to determine transformative.

A derivative work is typically what the generative AI produces where you could see elements of an original work but there are changes (i.e. the source work but shaded purple).

A transformative work may be literally the source work but used in a different context/purpose (i.e. a 5s clip of a 2h soccer game used to show proper throw-in technique as opposed to the original purpose of a sports game).

In the big picture, I expect generative AI to replace stock imagery and in this area I think this court case throws a huge wrench into that. If you're licensing generative AI instead of licensing the source stock imagery this case is going to be cited as why what you've done is illegal.

[1]: https://news.ycombinator.com/item?id=35991914

I take issue with “generative ai produces where you see elements of an original work but there are changes”. Generative AI does not take the pixels of the original work and collage them together. You have go out of your way to tell something like SD to take the input picture and shade it purple. Generative AI isn’t even the right tool for the job as that’s better done in photoshop.

Because this case stands for the proposition that when you make a derivative work, you cannot use it in the same way the original was used, for the same purpose, and call it transformative. The transformative prong of the fair use analysis is one of the most crucial prongs (if not THE most).

Right, but if you do sell your output, it's going to be bad news. I imagine most generative services intend to make a profit at some point.

You were doing so well at the start, and then you had to generalize it in a stupid way.