My guess?
Lots of pictures where different numbers of fingers are visible. If someone is holding something from behind with two hands, and only the fingers are visible, that would look like more than five.
The below could easily read as ten fingers and no palms.
I like to think of it more like biometric markers. They are definitely processing other peoples art, pulling from it various parameters - color choices, style. composition, etc. They are storing that information which is then used to generate the artwork. So I would say that this is more similar to, I dunno…the Blurred Lines lawsuit, which was claiming infringement based on copying a style, as opposed to sampling. And won.
"“I did that in ‘Blurred Lines’ and got myself in trouble,” Williams said. “I really made it feel so much like (‘Got To Give It Up’), that people were like, ‘Oh, I hear the same thing.’””
As for the difference between what the machine learning is doing and what people do, it’s a matter of industrial scale. A person can go out and study art, learn about a limited number of other artists styles and techniques, and then use that knowledge to create a relatively limited number of works in their lifetime.
A machine learning program can take in millions? BILLIONS? of pieces of art, more than any human ever could, and then spit out, in seconds each, millions or billions of images using that base (and it could be running on many, many systems while doing this). Not at all the same as what people do. In addition, a human artist will usually develop their own style, rather than always cribbing someone else’s.
So, I don’t think there is a real parallel here. What the machines do is completely different from how people learn and create art. So, we need courts and legislators to wade in
(This is not all directed to @cepheus42 just a convenient place to reply - I think we are in agreement on this at the high level)
To be fair - human artists often struggle with hands and feet!
Ha! True–but the human knows that they’re drawing hands and feet (badly). These AI things have zero consciousness–they have no idea what a “human” is let alone a specific body part. They really are just… cut, copy, and paste guided by updating sets of probabilities.
If I wrote and published an novelization of the new Avatar movie, I would absolutely be sued for copyright infringement, even though all the elements I used were, individually, uncopyrightable (e.g. character names, plot points). Every copyrightable thing is made up of smaller, uncopyrightable parts.
You can technically be sued for anything. Whether you win or lose is something different. What you’re describing is a trademark violation, not a copyright violation. Sure, Disney probably would sue for copyright infringement and a jury might be impressed by Disney’s expensive lawyers, but that’s not a strong argument.
To argue that it is not, the defendants would need to argue that the model is transformative fair use, not derivative. This might seem an easy claim. After all, an art generating program is very different from a picture, right?
They don’t have to argue that at all, actually. They only have to argue that the use of the works in training the model is fair use. The model itself doesn’t use the original works when generating a new image.
All four factors of fair use are on the side of the artists here.
The problem is that you’re weighing the result against the four factors, not the training, which is the only part where the original works are used.
- The nature of the use is commercial, and it is not transformative in the sense that it “adds something new, with a further purpose or different character, altering the first with new expression, meaning or message.”
The use isn’t all commercial. Stable Diffusion can be downloaded and run on your own computer for free. The result is absolutely transformative. It creates new images that no one has ever rendered before.
- The entire image is used in the production of the new work.
Except it’s not. The entire image is used in training the model and no part of any image is used in the production of the new work. The entire image would have to be present in the model and none of them are.
- The effect on the market is obviously bad for the copyrighted work.
Not necessarily. Some of the copyrighted works used in the training weren’t being sold, so the market effect would be negligible. But you also can’t reproduce the original works well from the models, so how would the market for that particular work be affected except through competition by different works? By that logic, nobody but certain movie studios can film action movies because other studios making action movies would affect the market for their films.
We also need to point out that this factor includes hypothetical derivative works. There’s no reason an artist couldn’t license their work for an ML model.
This is literally the only viable argument that could be made right now. If an artist were already licensing their work for use in AI models and someone else used them without the license, then they’d actually have an argument, but even though, it hasn’t been determined if training on an AI model is a fair use or not, so that might be a right not granted to the artist by copyright law. You’d have a better chance arguing this in Europe with moral rights involved in copyright laws.
Oh yeah, for sure (though it is, I believe well above the pixel level that the analysis is happening - they are getting markers for many things within the image beyond at the pixel level)
But of course, it doesn’t understand any of it, so you get weird hands and feet, both for the reason you pointed out (sometimes hands and feet in photos look weird) and the one I mentioned (trained on a lot of bad, though human drawn, hands and feet)!
The single image training program experiment was meant to show whether or not the original works are somehow encoded into the program. Or if the program only contains some sort of metadata thst does not fully represent the original image.
If a single image trained AI would reproduce only the original image, then it still means that a meaningfully complete version of the original image is stored in the program’s code.
As more images are added it might be harder for a user to discern which parts of which images are being used, but the entirety of each image would still be encoded.
We already have licenses that cover that exact situation. If you cover a song you have to pay a license fee. In this case the original artist has no choice, it’s a mandatory license (though most humans will still ask for permission). And, in fact, that license does NOT allow you a copyright if modify the song to be an (slightly) original work. The classic example of this is Jonathan Colton’s cover of Baby Got Back. He did not just cover it, he also changed the arrangement and the lyrics. Later Glee literally stole his arrangement and the lyric change he made that included his name ("Jonny C’s in trouble). He had no case to do anything about it though because his arrangement and lyrics changes were not covered in this situation, so Glee could get away with it.
Copyright is far more complicated than you seem to be claiming and I think that the status of machine learning art and text training and output are far from clear at this point.
It is absolutely a copyright infringement. Just like making a film adaptation of a novel under copyright would be infringing, regardless of trademarks. Adaptation rights are a very old part of the copyright bundle.
The training program and the model to generate new art are separate, copyrightable programs. The latter is the result of a combination of the the former and the copyrighted art.
Open AI sells credits to generate images. Some programs do not, that’s true.
As for transformative, I think you’re confusing a colloquial and legal definition. A non-human is legally incapable of having purpose, or intended expression, meaning, etc. If the model is found not to be infringing, this would be a good argument that the copyright should go to the person who selected the prompts.
The “new work” here is not the artwork generated by the end user, but the model that results from the initial art selection and the Machine Learning program. (Note: whether this creates a new program is technically moot, as RAM doctrine applies).
Once again, you’re conflating colloquial definitions with specific case law. To give a specific example, Rogers v. Koons, a sculptor made sculptures of a photographer’s photo. The photographer did not offer sculptures for sale, and had no plans to do so. The potential market for such sculptures was proven by the sculptor’s sales. The court found in the photographer’s favor.
More broadly, it’s self evident how generating art in the style of particular artists, especially those who are open to commission, hurts those artists’ sales.
there’s apparently some claims that you can transform a net into a decision tree. which then makes them ( in theory ) analyzable. as more research dollars begin to flow, it probably won’t remain a black box forever.
Given the size of the models…would a fully trained model ever actually be understandable to a human being? If only another computer can understand it I think it’s still a black box. I don’t know enough to say
ETA:
The opposite has also been true, and I think it’s applicable here. You can’t publish a picture of The Bean in Chicago without getting hit with a claim by the sculptor. This matters because the photo of The Bean is not The Bean, it’s a two-dimensional representation of The Bean…just as what a machine learning system stores is a mathematical representation of the art it is fed, not the art.
Based on this logic the representation of an artwork stored in a machine learning dataset could very well be an infringement even though it does not include the original art itself.
(Note that I think the whole thing with The Bean is stupid. If you don’t want people taking pictures of your art don’t put it in a fucking public space. But that doesn’t matter here)
While I don’t rate the claimants’ chances, it sounds to me like they have understood the situation rather better than a lot of people by not getting hung up on hazy arguments about the technical details.
100% of these models’ output comes from the total set of images they have consumed. If they owe everything to those images’ creators as a population, it is reasonable to ask how they can at the same time owe nothing to any individual artist.
It’s irrelevant what processing the machine does, or whether or not its programmers know how to audit that process. If I put a hundred people in a machine and eat just one sausage link from the other end, you don’t need to know exactly who I’m eating to know whether I’m a cannibal.
I would liken it to the “dilution” defence used by pre-1970s polluters: “sure, chlorobenzene is bad, but the ocean is big, so a couple tons spreads out to basically nothing”. And I think the harm in question is similar, too. In most cases, AI won’t hurt specific artists; it will taint everyone’s cultural diet, and future artists will be a bit worse because there’s a slight oily film over the culture they’re raised on. By the time we find out whether there is such a thing as esthetic mesothelioma, it will be late to do anything about it.
We already have licenses that cover that exact situation. If you cover a song you have to pay a license fee.
But a cover isn’t considered sampling.
And, in fact, that license does NOT allow you a copyright if modify the song to be an (slightly) original work.
And that’s irrelevant because the original work isn’t being copied or modified in Stable Diffusion.
Copyright is far more complicated than you seem to be claiming and I think that the status of machine learning art and text training and output are far from clear at this point.
Copyright is far more complicated than it’s possible for us to ever cover in a comment section. I would never claim otherwise.
It is relevant, at least in regards to whether the resulting program is copyrightable. There’s no standard judicial test for exactly how much AI makes something uncopyrightable, but the less control (or even understanding) you have of the process the worse your argument is for the necessary originality
I’m a paralegal who works in IP. I’m not an expert in copyright (I mostly work with trademark), but I’m not just talking out of my ass
it’s not clear to me either. it’s kind of like, is a cluster of stars or a bunch of neurons understandable?
i think being able to transform it into something like a decision tree means you’d be able to do statistical analysis, compare different trees of different models, and maybe in some cases look at particular paths to analyze a decision made…
is that useful? maybe? it’s all quite far beyond me
You weren’t talking about sampling. You were talking about playing a song by ear. Hence my entire post.
ETA: I was just pointing out that no, it’s not considered sampling, it’s something else that we already have laws and rulings for, so it’s not a good argument unless you think that the companies creating the datasets should have access to a mandatory license that they pay to each creator whose work goes into the model.
Which I am ok with
For a long time I did not understand how fanfic could infringe on copyright. I mean, it’s drilled over and over that copyright doesn’t protect ideas, it protects specific works. But there’s that case about the unauthorized movie sequel which, same thing, it’s not an adaptation, it’s not a sequel, and it’s not a copy.
It wasn’t until recently that I ran across the so-obvious-I-feel-stupid explanation: A translation of a book into another language will likely share none of the words. An adaptation from book to movie will likely be missing everything that isn’t dialogue, and even that may be greatly altered. They’re not the same work at all. And yet, copyright covers derivative works.
Obviously then, the tests that determine whether a work is derivative can’t depend on how much is copied from the original work exactly. It has to be about the plot, the characters, etc… which means even fanfic and unauthorized sequels and such can be identified as derivative works.
I hate it, but I accept it.
That said, I’m pretty sure you’re wrong about the model being a derivative work
If I generate a frequency list of English words by processing the contents of a library, I both could not have created that list without the work of every one of those authors, and yet I owe nothing to any one of them. Because words are not individually copyrightable, and in tabulating them I severed them from their context.
Indeed, some frequently used stock image backgrounds - used in thousands of distinct images (or more?) - were essentially memorized by the model.
However, that really only helps Getty Images… I doubt it will help this lawsuit.
My favorite example is when the USPS got sued for copyright infringement because of a stamp with what they thought was a picture of the Statue of Liberty, but was actually a picture of the copyrighted Statue of Liberty at NYNY hotel in Las Vegas.
The USPS lost
Sort of. Copyright is a bundle of rights. The right to copy is the main one, but translation, adaptation, derivative works etc are all included in that bundle and can be owned and sold separately.
