Judge lets art trio take another crack at suing AI devs over copyright A US federal judge almost threw out a lawsuit brought by artists accusing text-to-image AI developers of copyright infringement, but decided to give the creatives a chance to improve and resubmit their complaint. In January, illustrators Sarah Andersen, Kelly McKernan, and Karla Ortiz sued Stability AI, Midjourney, and …
If I, as a human, read a book by Author X, then write a book "in the style" of Author X (yet do not do a 1-for-1 copy of Author X's book), I'm legal in doing so, and owe Author X nothing, right?
What have I done here? I have: (1) "trained" myself using Author X's work, and, (2) produced a work derivative of Author X's style, and sold it for profit.
Is Author X legally entitled to a cut of my profits? I think not.
Does an author who has written a book and had it published have any control over what uses other people put that book and its contents to? Nope.
Does a publisher who employs me to write a book in the style of Author X have a legal requirement to pay Author X a portion of the publisher's profits on the book I wrote in the style of Author X? I think not.
Given that, why do the creative types think the publisher owes them money, just because the publisher used a computer to do something a person could do without incurring any legal obligation to Author X?
(Icon for Devil's Advocate)
Just to respond to one point of your comment.
Right at the start, you wrote if you read a book by Author X. Assuming you have not pirated that book, then you have purchased a copy of that book. Either directly by buying the book, or indirectly through the licences purchased by libraries to lend you that book. Hence the Author X is being paid for their work.
If you read every one of their books, then you would be paying each time for that book.
When the AI scrapes all of those books off the internet, has it paid for them? No. Not one cent. If the AI firms had paid for a single copy of each book they scraped, there would be no question here, no lawsuit. But they havent and naturally, they dont want to, because paying for every single one of the millions of books they scraped would put them out of business. And I would hardly call that fair use.
If the judge rules that the AI firms have to pay the cost of purchasing one of each book they scraped, I would suggest that would be fair. But you can bet the firms will fight even that, as it would open up the door to them being on the hook for millions (if not billions!)...
@lglethal: you make an excellent point. I wrote presuming "I" had either purchased Author X's book, or borrowed it from the library.
I guess the creative types are back to chasing after the owners of websites which illegally reproduce (make available for viewing and/or download) those creative types' works.
There is copying and there is copying. A spoof of say the Lord of the Rings is fine if it changes names to sillier ones and rewrites the text sufficiently to avoid direct plagiarism of stretches of writing. But if you copy-paste blocks of text into your own draft and follow up with a little find/replace, then you will be nailed.
I can draw a cartoon of a comic dog sitting on a roof and yelling abuse at an old aeroplane. But if I trace Schultz's work, call the dog "Snoopy" or have it yell "Curse you, Red Baron!" then I am nailed. Or, if I give my dog a flying helmet, make the machine a triplane and have it yell something else, then I have to have pockets deep enough to feed a generation of respected legal professionals.
I cannot imagine that the law will treat generative AIs any differently. The burden is on the plaintiffs to demonstrate plagiarism rather than mere reference. Frankly I don't give 'em a cat' s chance in a furnace, but what do I know. The cynic in me suggests that the judge either gave them more time to please eloquent legal council making "kerching!" arm movements to each other, and/or because he hasn't got a clue about generative AI.
"A spoof of say the Lord of the Rings is fine if it changes names to sillier ones and rewrites the text sufficiently to avoid direct plagiarism of stretches of writing."
You might find the guardians of Tolkien's estate taking issue with you. They (or possibly it was the film company, I can't remember which) have been know to come down on what they considered misuse.
Upvoted you, and I have read and found mildly amusing "Bored of the Rings" (especially "stay out of the kitchen") but to put the other point of view JK Rowling sued (and I think won) an author because the book focused on a magical child going to a magic school.
JK Rowling sued (and I think won) an author because the book focused on a magical child going to a magic school
Because she invented that concept in 1997, and novels such as The Little Broomstick (1971) and The Book of Atrix Wolfe (1995) never existed.
I'll admit Rowling's version is far worse, so maybe she has some grounds for complaint there.
While unfamiliar with the cases, I looked up a summary. The cases I can find where a suit was successful involved using the characters' names at least and often large parts of the setting. So it's not just some child being sent to a magical school, but children using the same names from being sent to schools invented for that book, sometimes both applied and sometimes only one. Of course, there may be others that Wikipedia didn't choose to list.
If you take the source code for Adobe Photoshop and produce your own binary based on that source code, is that copying?
Is it still copying if you use a different compiler from the one Adobe uses?
Or if you write your own compiler?
If you take a work written in English rather than C, and use your own compiler to produce a binary, does that make any difference?
A problematic comparison - when using a programming language compiler we are deliberately and knowingly invoking a series of transforms that are explicitly intended to create a copy of the meaning of the input source code. Compiler writers take pains to record how the input has been copied (via, for any given compiler, a fixed set of well understood transforms) into the output - we use this record in our debuggers to demonstrate that the copying has occurred and to determine how it has been modified - i.e. optimisations such as constant folding, loop unrolling.
We also have an agreement in the profession that these automated transforms do not constitute making substantive changes to the original work. This has been thrashed out over the years, despite attempts by compiler writers in the past to make claims that they owned the output object code and therefore they could put limitations on what you could do with it (we see remnants of that in licences that cover, e.g., "if you make more than $XXXX you must pay more for your licence").
In a compiler, the intent is always to maintain a 1:1 relationship with that the input source expresses and the output binary performs - without that, we just call the compiler a broken, useless thing. The uses of anything we call a "transform" are intended to maintain that relationship - in effect, the selected transforms are "trivial"[1] and comparable (in lay term) to just shifting the axes (and even that we'd expect to demonstrate to the layity with simple pictures rather than just using those words).
> If you take the source code for Adobe Photoshop and produce your own binary based on that source code, is that copying?
Yes, by the agreed-upon terms common to our industry
> Is it still copying if you use a different compiler from the one Adobe uses?
Yes, see above.
> Or if you write your own compiler?
Yes, see above.
> If you take a work written in English rather than C, and use your own compiler to produce a binary, does that make any difference?
If - and only if - the English was a precise and unambiguous description of how to compute some value (aka an algorithm) *AND* your use of the word "binary" is to be taken with same meaning as for the previous questions, then - No. No difference. But then again, to get that level of precision you must have been selective over what form and content of "English" you'll accept: congratulations, you have just invented COBOL-2023.
HOWEVER if your last question was intended to refer to the use of English and the behaviour of the LLMs - where the transforms are *not* intended to maintain a 1:1 relationship (as evidenced in the ability to use phrases like "in the style of"), where there is, as yet, no tried and tested agreement within the professional community that this relationship is intended to hold, and, if it ever doesn't, the tool is broken, where the word "binary" is being used to describe something other than something that is executable (in whole or following linkage) - then, in that case, the only response to your question is:
Mu.
[1] Not trvial to encode or get right, not even to think up in the first place, but in "artistic" terms compared to the sort "transformative" acts that copyright usually gets involved in: for a start, most (all?) of those are lossy whilst ours are even capable to increasing precision - e.g. type inference.
By English, I mean a typical book that you might find in a book store.
By binary, I mean an large language model of the sort you can download from HuggingFace.
The compiler whatever is used to turn your collection of books into the large language model.
Then the answer is: mu.
Your first three sentences describe one situation.
Your last is, as you confirmed, describing a totally different situation (as outlined in my previous post), so the question "does that make any difference?" is malformed, hence: mu.
Whether you can legally copy and compile *ANY* source code depends solely on the license that the author(s) applied to it.
If no licensing details are included in the source code, then you can do what you like.
If the code was in a book , e.g. "Software Tools in Pascal" (Kernighan & Plauger) or Sedgewick's "Algorithms", the publisher's copyright notice controls what you can do with the physical book and its content's representation on the page.
In both these books the copyright notice says explicitly that than you can't reproduce any part of the book. However, as all the example code in both books is written in Pascal, I, and doubtless almost all other software authors, have always assumed that translating the Pascal example code into another language, i.e. C, Java, COBOL or even assembler counts as 'fair use'. In both books the 'Preface' text also makes it clear that the authors expect the example code to be copied, possibly translated into a different language, and incorporated into other programs and then compiled and used within other personal or commercial programs. To me this means that the publisher's copyright notice is meant to apply just to the physical book and the text it contains as represented on its pages.
> publisher's copyright notice is meant to apply just to the physical book and the text it contains as represented on its pages.
Yes (the exact contents of the book). The "in part or whole" (or similar) is subject to conditions in your jurisdiction (e.g. you may have various "fair use" rights).
> If no licensing details are included in the source code, then you can do what you like.
Sadly no. Short sections may fall under "fair use" but longer sections, nope.
Teaching books and "cookbooks" nowadays include explicit code licence terms, because everyone is now aware of the issues and the use of informal notes in the preface weren't really sufficient.
> If no licensing details are included in the source code, then you can do what you like.
Oops, misread your post: when you wrote the above you weren't, as I'd thought, talking about source code in a book (it is only later on that you mention a book).
If you just come across some source code - in a random file somewhere, in a printout lying on a park bench - and there are no licensing details included in it STAY WELL AWAY FROM IT! It probably isn't worth the risk.
It will still be under copyright, even if that is not explicitly stated, even if no author is named[1] . Aside from anything else, you may not have a complete copy of the original[2].
Although you could get away with reading the found source code, you almost certainly don't have the right to input it it, let alone compile it at all (although, as usual, you can generally get away with it in the privacy of your own home), let alone "do what you like".
[1] no, you can't assume "Public Domain" - aside from the point that PD is only a US concept, without knowing the authorship you *potentially* have an "orphaned work" which is a whole other can of worms - keep away, really not worth the risk; you'll have to fight Google[3] for starters (in the US) :-)
[2] if you rip apart a novel (say, "Whispers Underground", Gollancz UK 2012) and walk off with the middle third, it likely[4] does not provide author, title or other copyright information but it is most certainly still under copyright (by Ben Aaronovitch in the example) all the same.
[3] Google’s Plan for Out-of-Print Books Is Challenged
[4] some novels have the title and author on each leaf: this is a book design choice. I've just checked a pile and none of the newer novels (in this sample) do it, but the 1970s books do.
I gave an upvote, but... It depends on the copyright the code is bound by - and that doesn't necessarily have to be in the header of each or any module of the code. If there is no header in the code describing which copyright model it is held under, you will need to find out which one applies. You can't just assume, because there is no copyright notice or open source license etc. in the code that it is totally public domain, you should assume that it is copyrighted, until you can prove otherwise - E.g. if it is internal code from a company that has been stolen, they might not have copyright headers, because they never expected that the code will never be seen by anyone outside the company...
The problem comes when we try to decide whether LMM processing counts as reading and understanding, like a human brain does, or mashing and reconstituting, like a dumb program does. It isn't clear, and although media hype likes to paint these programs as sentient systems that build an understanding of concepts, their output indicates that they are not.
Parallels are difficult. Here's an attempt. I have recently opened up an embedded software image and extracted the important program from it, which I am disassembling. It's not even clear whether I'm legally allowed to do that, which is one reason I won't name or describe the product. I'm interested to see what this software does and I'm just reading for that point. However, let's assume that I was going to use that software for my own purposes. If I were to port that software to run somewhere else, large parts would be left behind. The entire UI system is related to the manufacturer's proprietary hardware, which I wouldn't be using. All that code would have to go and be replaced by something that could handle a different set of interfaces. There are some components I could do without. Those would also be deleted and the various gaps covered. There are also certain aspects where the software could use some improvement, so I'd be making lots of edits. Since I don't have the original source code, a lot of the decompiled bits would be significantly changed, by compilers and other tools, in order to fit them into my modifications. My theoretical version would end up being very different, both in appearance and in function, to their original version. If I published it, there's a reasonable chance they would never notice that I had used their software in developing mine. If they did figure that out, especially if I admitted what I had done, they would not accept that as legal use of the software and neither would a court. It didn't matter that I've used only a small part of their work and added plenty of my own, it wouldn't count that my binary didn't contain chunks of theirs, and it wouldn't matter that software made many of the edits; that wasn't my software to use.
Bots that cheerfully quote sections of text or reproduce original images make it rather clear that they have no ability to avoid printing significant portions of copyrighted content. There's an argument that, even if the program doesn't print the content, it is not legal for it to be trained on that content unless it has been legally obtained, and since copyright does not include an automatic license to produce derivative works, it may not have been. This won't be the last time the human-to-bot analogy will be made. We will have to decide for ourselves whether having a computer provide plausible deniability by randomizing things so much that you might not get the original input back is sufficient to equate its actions with that of a human. I don't think the legal arguments clearly indicate a winner.
Oh, if only it were all so clear cut and simple :-)
One could happily agree with what you say and at the same time start to poke holes in it: as you so rightly say, parallels are difficult.
As usual, the lawyers and their twisty minds are going to be the obvious winners. Well, a small number of them, at least.
The difference here is you’ll need up to a year writing that book, while AI will need seconds, flooding the market with copies of someone’s style, for which he or she needed up to 20 years to develop. You see the catastrophe here, this type of stuff will kill this type of creativity in humans, because it won’t be worth it, if a bot can churn out “inspired” copies a thousand times faster.
I just recalled that many books have provisos in their front matter forbidding mechanical reproduction, etc., typically as relates to copyright, and, some of those may have used the words "scan" and/or "scanning" in their list of prohibitions. Ingestion by AI/ML systems might fall afoul of these; whether some or all of those provisos are legally-enforcable probably will require a number of additional court cases.
(Icon for solicitors lighting their cigars with large-denomination paper currency as they laugh their way to the bank.)
I hold no love for lawyers but what they do is argue cases, often ambiguous natures, then from that we can draw precedents that help all of us in future to know if we have a case or not. So while I'm all for a world without lawyers, I'm afriad that despite their high salaries, they do provide something worthwhile that could benefit most of us at some point when we need it.
"So while I'm all for a world without lawyers... "
That's not even a sensible wish. Every society has/has had lawyers. Sometimes it is just one person (usually a man), such as a tribal chief, or shaman/medicine man/seer; sometimes it is a bunch of the leader's mates; sometimes (rarely) it is a profession with standards and the ability for members to be held to account by the public. Whichever model you choose, lawyers are a) powerful, b) influential and c) rich compared to other members of that society. By and large, we have the final category*, so which other model would you choose (not confining yourself to my examples)?
* I am not arguing that the current system is the best, but that it is one of the best so far.
That way leads to murky waters.
Want to bet that someone in a Judge's gear is going to end up taking terms that referred to older technologies (mechanical reproduction => printer's press, brass rubbing, newspaper burnishing; scanning => photo litho, fax), re-apply them to computers and end up banning the sale of ebooks and ereaders in their jurisdiction!
Not a problem, since that is already how the law works. If I scan a book, make an ebook out of it, and sell that ebook, I'm violating copyright. If the publisher does it, they have the copyright and can grant themselves the rights necessary to make that ebook. The law already works as you describe, and it always has done. In my opinion, that's how it should work.
This post has been deleted by its author
> Want to bet that someone in a Judge's gear is going to end up taking terms that referred to older technologies, re-apply them to computers and end up banning the sale of ebooks and ereaders in their jurisdiction!
>> The law already works as you describe, and it always has done
No, nothing about making illicit copies (don't know how you read that into the previous comment)!
What was suggested was TOTALLY BANNING Ebooks, banning the entire *concept* of ebooks as a legal means of distribution.
As in, the Judge getting the whole thing so messed up that they fall into (hopefully/probably) unintended consequences!
Since the law works on the basis that copying is not allowed unless you have the right to, it already holds that ebooks are illegal unless you have permission to make them. Interpreting it in a different way would be very difficult. For example, even if a judge takes the strongest position against the creators and operators of LMM programs, giving the most rights to the original creators of training data, it would not prevent them from making something out of training data to which they have the copyright. Microsoft could still train on the code they wrote. An author could train on their own writing (subject to their contract with their publisher). So if you really think a judge could somehow ban ebooks as a concept, I'll take the other side, and I'm curious to see your proposed logical error that could lead them to do so and have any legal viability (if it doesn't have that, then another court will quickly throw it out, so that will be a requirement to obtain the situation you describe).
...And from my pint of view, yes, the artists have a case. Proving it will be hard however. It all comes down to the model being used and the prompt given to it.
Prompting "By Rene Magritte" and "Inspired by Rene Magritte" gives two different images. It gets even more obscure when you combine artists as you try to create something meaningful and unique. That is the true power of SD as far as I am concerned. Not that it replicates what one artist did in the past but that it combines the styles in meaningful ways.
My "artwork" has even been promoted by StabilityAI in their blog because I submitted it on their Discord channel. All I did was find the right combination of words to create something new.
So if I am prompting using multiple artists am I violating copyright laws?
> Prompting "By Rene Magritte" and "Inspired by Rene Magritte" gives two different images.
I mean, that makes sense? People who paint images called "$foo: inspired by Rene Magritte" are probably a style of their own.
We shouldn't think of the prompt as expressing an intent, but rather as selecting a stylistic reference class.
Recalling past conversations with other SD users I have to say that some of them only want to work with models that can replicate a given artists style. They are not interested in the the interpretations of the AI...just the regurgitations. In my mind this is due to a lack of imagination on their part but I cannot speak for them because I am not telepathic. However the overwhelming number of anime pics by artist "*" that I have scrolled though is staggering. So much regurgitation that I stopped posting my results just so I wouldn't have to see it anymore.
>> So if I am prompting using multiple artists am I violating copyright laws?
If it works anything like copyright in the music industry, where one "artist" (using that term loosely) samples tracks from other artists to create something "new". but is just 6 songs mashed together, then the artists seem to be out of luck.
> " ... Until now, when a purchaser seeks a new image 'in the style' of a given artist, they must pay to commission or license an original image from that artist," the trio's complaint [PDF] read.
So if I want a picture "in the style of Picasso", there is just nothing that can be done. As Picasso is dead, this request cannot be satisfied.
Kind of sad how each artist has their own perfectly unique individual style that disappears with their death and cannot be resummoned or imitated by any means known to man.
Saying they need to provide more facts as it seems "implausible" is evidence of the judge not knowing how this AI tech works.
Which kinda makes sense, in general judges aren't the most tech savvy bunch. It's down to the people suing to outline exactly how the system works.
There have been judges who have made the effort to become familiar with the arena of litigation - William Alsup in California comes to mind (Oracle Vs Google). However I wonder if some think that may be counterproductive in an essentially legal judgement.
There is a need to show that it didn't pick up "in the style of" from some other source - like wannabes posting their attempts at "in the style of", parodies, rewordings etc. Not forgetting all the repostings of favourites on Twitter, Facebook.
Not that these are good arguments, but you have to prepare for what the defence will throw at you.
Not commenting on the merits of the case, but to some extent, it's the job of people bringing a lawsuit to be able to explain the problem to the judge, or at least enough to convince the Judge there's a problem to dig into.
The judge can do research later, but step 1 of most cases is proving to them that they need to research *this* , and not move onto the rest of their caseload.
"Saying they need to provide more facts as it seems "implausible" is evidence of the judge not knowing how this AI tech works."
That's not how I interpreted that bit. I thought it meant that the judge thought it implausible merely on the evidence presented. Hence, giving them the chance to go back and come up with a better case.
"Until now, when a purchaser seeks a new image 'in the style' of a given artist, they must pay to commission or license an original image from that artist,"
Really?
Because a lot of webcomic artists, such as the aforementioned Sara Anderson, have incredibly simplistic styles, and hiring a different artist to draw in that style wouldn't be hard at all.
Moreover, if it's possible to 'own' a style, then it becomes necessary to ask what defines that style, and whether other people can use those design elements without infringing on the style.
Does Randal Monroe now own stick figures?
Does Yves Klein own the color blue?
In fact, we already have a body of law about defining visual elements that designate the origin of a thing. That's what trademarks are about. If someone creates an A.I. image with mickey mouse in it, they'll be unable to use it commercially. If these artists really want claim they're entitled to protection automatically, they'll need to convince me that trademarks are automatic and non-specific, and I don't buy it. If it was possible to trademark Walt Disney's style, the house of mouse would have done it years ago.
There might be an argument to be made on the use of the images; if they've licensed the image for use by humans and not computers, then scraping and training with it could be prohibited. But that's input, not output; as to the companies owing them for images made using their 'style'? That way is a slippery slope, and I don't think the judge is insane enough to try that.
That's not what they are claiming or asking for. They are not saying that they own a style and you can't produce anything that looks too close to that. They are saying that the bots are only able to imitate that style because they have, without permission, used a lot of their copyrighted artwork to produce those new pictures. In some cases, identifiable pieces of that work have been produced by models such as this. The court has to decide whether those parts are sufficient to be violations on their own, and whether there is a right to use copyrighted content as input to a program without permission and despite complaints from the copyright holders.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
