The image synthesis lawsuits cometh
Have you ever sued Stable Diffusion at midnight, Batman?
The very same lawyers that sued Github Copilot now filed a lawsuit against Stability (Stable Diffusion), Midjourney and DeviantArt “for their use of Stable Diffusion, a 21st-century collage tool that remixes the copyrighted works of millions of artists whose work was used as training data.”
AI-Copyright-expert Alex J. Champandard has two good threads about the trial, and even if he thinks that the lawyers present a sloppy constructed case, he also says that Stable Diffusion, Midjourney and DeviantArt are in a bad position.
The classic response to the lawsuit in the SD-Reddit-sub have been jokes about how the website introducing the lawsuit portraits Image Synthesis as a "collage tool". Which is… fine, because no it's not a tool for classic collage. But also, it doesn't matter much.
That's not because the website for the trial is not just saying 'its a collage tool', but a '21st century collage tool', which is something else and conveys the innovative nature of the ai-system. It's because a recent paper has proven that image synth tools can reproduce training data verbatim. All Stable Diffusion-Checkpoint-files "know" what Batman looks like, because Stability and LAION used tons of Batman-images, without paying a dime to Warner, and sometimes it puts out that data unchanged. That's all the lawyers and their plaintiffs need.
Also, this lawyer has a very big case going on and he will never under any circumstances say that “yeah sure, this may not be a collage tool, it is actually a complicated machine that is able to generate anything on prompt, and the outcome is kind of random so you never know” or whatever. It’s a lawyer, he wants to win a case. He will say: "I can produce a billion plausible Batmen, and thus, this is a collage tool".
Not because he knows better and is a liar, but because it’s true and it’s his job to present his case in a way that will most likely win over the opinion of the court. And I think he has a good chance it will, because even when the court follows the more complex description of image synthesis as “a tool for denoising an image based on parameters that makes up its latent space”, those parameters still where developed by a for-profit company incorporating the works of thousands of artists without compensation. I'm absolutely not convinced that a Fair Use-defense will apply here.
The question then boils down to copyright for styles and it is by no means save to say that the law doesn't protect styles (yet), because you simply have to introduce a new law which does. The question is if such a law would be fair and just. And maybe, in world where i can make a fast snapshot of your drawing and generate whole worlds in your style, it is. (And style is not the only copyright issue of image synthesis, but trademarks too. Imagine a pen that "knows" what Batman looks like, but hasn't been developed under the license of DC Comics. In my basic understanding of copyright law, this would be a crystal clear violation.)
I'm very sure that if the emergence of AI-image synthesis leads to new copyright laws involving style, those will be limited to automization systems and AI, and you will still be free to go and copy Picassos cubism by hand or in Photoshop, and you will also be free to train an algorithm that "learns" styles and mimic them, but you will not be free to publish that algorithm and it’s database of weights.
Just as most of fan-art today is illegal by law and merely tolerated by IP holders, AI-algorithms will join that rank. You will always be free to draw batman, and scan one billion batman comicbooks and train any technology with those scans and create the most sophisticated “intelligent”" Batman-drawing machine ever, but as soon as you publish this output or the algorithm, you can be held accountable, just as you can be held accountable if you print T-Shirts from Batman-drawings in your private blackbook.
Artstation and Deviantart will do some TOS-shuffling, and art/illu/photo-platforms will develop compensation models akin to Shutterstock, who want to pay their userbase of photographers for licensing their images to training datasets, most likely to build their own stock photo image generators.
Stuff like Have I been trained by Mat Dryhurst and Holly Herndon will morph into a consent-mechanism for image platforms. Stability and LAION already said they will respect those mechanisms. With those new consent mechanisms and the treasure trove of creative data, those platforms can train legal CKPT-files for creative applications, and the artists involved in the training data will earn some money from that. And make no mistake: The illegal checkpoint-files from the black market, ready for download on TPB, will be better at producing stylistic immitations, based on tons of images coming from "illegally" scanned artworks, comics and illustrations. The only thing i'm really sure about is that this tech will not go away.
Another common argument you hear from AI artists is that those algorithms are just learning like humans do and that if you make it illegal to see an image and extract the "rules of style", any human artist could be held liable and all human culture basically would be illegal. This is shortsighted: No human has a corporate organization acting inbetween the perceived artwork and the artistic output. Humans see something, extract the rules of style, and then you apply those rules in your head, transform that into motorfunctions and put out lines on paper, from which an image of your mom emerges.
With AI its like: You want an image of your mom in the style of Picasso. The machine has “learned” what "picasso-likeness" is and has a parameter for that. It “knows” how a “mom” looks like, and combines those two, flattening a canvas of noise into what its statistical model says your mom would look like, if she was painted by Picasso.
Note that, with image synthesis, there is zero symbolic reasoning going on, the ability of humans to take any input and transform it into symbols and play with those, in more or less sophisticated ways. Image Synthesis doesn’t do any of that. It takes an input and maps it onto its statistical model and adds a stochastic element with its denoising-function. In human terms, those models haven't learned anything at all, just like a phone book hasn't learned anything about the people whose names are printed on it’s pages. This is why no algorithm is creative, and why they are called "stochastic parrots". This, too, is relevant in this case, because those machines don’t engage in any creativity that could be protected. There is no "learning" going on. This is simply about "machines complex enough to extract 'rules of style' from very big sets of data".
Also, there is an ethical dimension to this that i don’t see much talk about. In a recent paper researchers laid down a framework for the ethics of “mimetic models” of “AI that acts like you”. Arguably, artists and illustrators put a lot of their own personality into their work and that their style also represents a good chunk of who they are, as a person. I don’t think this is a privacy violation as we define it today, but any good lawyer could build another argument from that.
None of this says that you can't create art with those statistical models. Ofcourse you can create art with this tool. You can create art with anything, even with a banana and duct tape or old stinking fat and a chair. But that's not relevant for this trial.
The only thing that matters here is if the companies involved used unlicensed artworks in the creation of a commercial tool, and if the Fair Use defense applies to a machine that can mimic artistic style.
They did, and we'll see.
update 2023-01-18: A day after i wrote this, Getty Images sued Stability for ‘unlawfully’ scraping millions of images. Spokespersons for Getty compare the case to the Napster situation and Spotify.
Previously:
- Die rechtlichen Fallstricke von AI-Bildgeneratoren
- Stable Diffusion is creative Napalm and I love it
- Data Laundering with the devil
- AI-pirates and large language warez