Judge bins AI copyright lawsuit against DeviantArt, Midjourney – Stability still in the mix A judge has dismissed copyright infringement claims against DeviantArt and Midjourney in the US – and has allowed a case against Stability AI to continue. In January, three artists - Sarah Andersen, Kelly McKernan, and Karla Ortiz - sued the three aforementioned startups, accusing the businesses of using people's copyrighted …
1. The only legal standard that makes any kind of sense is, "if it would be illegal for a human to do it, it's also illegal for a computer system to do it, otherwise it's fine".
2. Nothing in copyright law allows rightsholders to place any restrictions on who or what can study and learn from their work. As long as they don't (copy, perform, distribute, adapt or translate) it, they're golden.
I'm sorry for the artists, but they're on a losing track here.
I can imagine if I was Sarah and I saw people giving prompts to LLMs like "draw me a cartoon in the style of Sarah Andersen about not wanting to get out of bed on a Sunday morning" and the model producing comic strips just like mine, I'd reach for a lawyer too. That's got to be some form of copying or adaption?
The problem is that copyright law only protects a person's specifc works, it doesn't protect their style, ideas or any other single aspect of the works. The law as it stands simply doesn't protect against what these generative algorithms actually do, and these plaintiffs would be better off campaigning for a change in copyright law than fighting this case on the existing law as it stands. And no, I'm not being naive - I know exactly how hard that campaign would be. That's how certain I am that the law isn't on their side in this case.
Just ask yourself... 'if I went to Amazon Turk and asked for someone to "draw me a cartoon in the style of Sarah Andersen about not wanting to get out of bed on a Sunday morning"' would it be copyright violation?' The answer is clearly: "it depends on whether the resulting cartoon is a copy of of an existing cartoon, or just in the same style".
And, like real-life copyright cases, in some cases the answer is no, others it is yes, and others "well, depends what you mean by a copy" - and a judge or a jury may go either way on it.
@Veti I agree with both your points 1 and 2. Yet I don't think they are on a losing track. Stability AI and LAION break point 2, both copy and distribute. LAION copies and then supplies (distribute) to Stability AI and Stability AI make copies to train their AI.
How is that any different from a human copying a book in a public library giving the copy to a teacher who makes copies as teaching material for their students.
It wasn't clear from the article if Stability AI or LAION actually supply copies of images or merely if they create some kind of metadata that they then pass on.
It is quite interesting to debate the difference between that activity and a human looking at an image and learning from it which is clearly what they are doing in practice.
> I am not convinced that copyright claims based [on] a derivative theory can survive absent 'substantial similarity' type allegations," he said.
The judge also seems to be saying that they would have to prove that the various systems output images that a reasonable person would think of as a copy by being substantially similar. I don't think people can claim copyright over original images "in the style of" some other artist, notwithstanding the legal issues with passing off of course.
I was writing neural networks to identify crops from satellite images in the 90s. Now while I don't know how the human brain learns, I would argue those networks were learning and so are the LLMs of today.
I would say that LLM are learning and that in itself is not copyright infringement and their output when it comes to copyright infringement should be judged no differently to the output of a human.
The copyright infringement is how the training data is obtained. In my opinion copying copyrighted work to train your AI and claiming it's ok because it's public, is no different than a school or University etc copying textbooks from a public library to teach their students and claiming its ok the books are public.
> In my opinion copying copyrighted work to train your AI and claiming it's ok because it's public, is no different than a school or University etc copying textbooks from a public library to teach their students and claiming its ok the books are public.
Your analogy has a redundant step. The school or university would be fine having their students go to the library and no copying is necessary. A better analogy would be for the lecturer to go to the library, read the book, then teach what he/she learned to students. Which is perfectly legal.
Even better: an art teacher studying public artwork, then teaching students the techniques to enable them to produce their own art. It would be somewhat derivative of what had come before which we expect because no art comes from a vacuum.
@ "Your analogy has a redundant step. The school or university would be fine having their students go to the library and no copying is necessary. A better analogy would be for the lecturer to go to the library, read the book, then teach what he/she learned to students. Which is perfectly legal."
The step is not redundant unless the LLMs are going to the internet and accessing the copyrighted material. My analogy is the LLMs are the students the internet the library and the training data copied from the internet is the book copied from the library.
How is that any different from a human copying a book in a public library giving the copy to a teacher who makes copies as teaching material for their students.
Can't speak for any other jurisdiction, but copyright law in Oz allows teachers to copy a single chapter or 10% of a book (whichever is greater) for use in the classroom.
Iirc, LAION is just a database of image URLs with various metadata concerning each image, and doesn't hold any copyrighted data directly. The only (extremely tenuous) allegation I can see making any sense here would not be against LAION directly, but instead that Stability created an unauthorised copy of the image at the point where they requested and downloaded the image for processing by the model's training algorithm.
The sticking point here though is that these requests are made over the public internet, and so you're opening a massive can of worms if you try to argue that receiving a copy of an image in response to a request, and temporarily retaining that copy only for strictly private use within the program that made the request, without further disseminating it in any way, is a breach of copyright. What's the difference between Stability doing that and, say, your own browser caching a Getty image from a news site?
The plaintiffs' lawyers in this case were attempting to side-step this by alleging that the training process is effectively making an incredibly highly compressed and somewhat lossy copy of the image, allowing it to be reproduced almost exactly by the AI, but the glaring hole in this argument is that they would need to be able to consistently prove significant similarity between output images and source data without relying on any kind of weirdness that Stability could argue as being reasonably unlikely behaviour by an end-user (e.g. brute-forcing a bunch of gibberish text prompts that just happen to produce training images by running the algorithm backwards). I think Stability are pretty confident in their ability to defend themselves against that tactic, and I don't see many other avenues that the plaintiffs could pursue if the courts reject that line of attack.
This post has been deleted by its author
"... but also the content generated by those models – incorporating themes, elements, and styles created by human artists – was unlawful derivative work."
So, we (human, or a supposed AI) are meant to admire an artist's work, but woe betide anyone seeking to learn from it, to improve upon it, to take forward nascent possibilities, to express excitement via emulation of something momentarily original which itself inevitably arises from a long line of 'derivation'? Also, note that should something stunningly original arise, that in any area to which anachronistic copyright applies, then restriction on derivation threatens to leave a shining jewel forever (at least decades) detached from the evolving body of human achievement.
No combination of legal phrases can result in a clear distinction between 'legitimate' derivation and that which infringes so-called copyright. There is one protection against losing the opportunity to raise income by sale of original copies of one's works, and only one, for which an unambiguous law can be framed to benefit an artist; that is against misrepresenting oneself as the artist (composer, writer, etc.) to gain pecuniary advantage. For example, signing a drawing falsely under the name of the true artist, and seeking to sell it. Works clearly resembling, or derived from, those of another person, could then avoid accusations of plagiarism if offered to the world under headings such as "In tribute to X", "Based upon a work by X", "A variation on X's depiction of ...", "A digital copy of ...", and so on.
"Dismissed all copyright violation claims made by McKernan and Ortiz since neither of them registered their work with the US Copyright Office"
I always thought the Berne Convention provided Copyright Protection at the instant of the work's creation, that there was no explicit need to register a work to secure Copyright Protection.
Is this just another case of 'Welcome to America'?
Several posts on this message board argue that scraping for training sets should be fine, because human learning works pretty much in the same way and that's legal.
I would point out that the lawsuit is not against the AI. The lawsuit is against the guys who made the AI and/or trained it. The thing they did is not the same as human learning at all, and it might not be legal.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
