I'm really confused by the FSF statement here. The court ruled that the use of copyrighted information is fair use. The issue is that Anthropic pirated (obtained illegally) copyrighted work and that was the offense. FSF books are free to download and store etc. The license says: "This is a free license allowing use of the work for any purpose without payment." So how can they claim that their rights were infringed when the court ruled that the problem was the illegal downloading of copyrighted work? It's impossible to illegally download a FSF book.
The issue is that every CS masters student & AI researcher knows how to build a SOTA LLM..
But, only a few companies have the resources.
The process:
(1) steal as much data from the internet as possible (data is everything)
(2) raise incomprehensible amounts of money
(3) find a location where you can take over the energy grid for training
(4) put a black box around it so nobody can see the weights
(5) charge users $$$ to use
(6) retrain models with user session data (opt in by default)
(7) peek around at how users are using, (maybe) change policies to stop them from using that way, and (maybe) rapidly develop features for that use case.
(Sorry that last one is jaded and not fair - just included to give you a picture of what could be happening with this sort of tech)
…
The entire premise of the product is “built on the backs of any & everyone who has ever published a work”
> It is a class action lawsuit… the parties agreed to settle instead of waiting for the trial…
It would be nice if members of the class could vote to force a case to trial. For the typical token settlement amount, I’m sure many would rather have the precedent-setting case instead.
If/when you get a postcard/spam email that you're included in a potential class action lawsuit settlement, you can opt out of the class (in which case you preserve your legal rights to sue separately) or file comments with the Court.
You can, but then you lose the power of a collective and have to manage a lawsuit yourself. If you are being represented as part of a group, then you should have means to direct that representation.
Surely some firms choose to hold referendums already, but I could see that being a good law! As Better Call Saul explored in its early seasons, the interests of the large law firm can easily diverge significantly from the interests of the plaintiffs.
The framing of 'share your weights freely' as a remedy is interesting but underspecified. The FSF's argument is essentially that training on copyrighted code without permission is infringement, and the remedy should be open weights. But open weights don't undo the infringement -- they just make a potentially infringing artifact publicly available. That's not how copyright remedies work. What they're actually asking for is more like a compulsory license, which Congress would have to create. The demand for open weights as a copyright remedy is a policy argument dressed up as a legal one.
In GPL cases for software, making the offending proprietary code publicly available under the GPL has been the usu outcome.
But whether you can actually be compelled to do that isn't well tested in court. Challenging that the GPL is enforcable in that way leads you down the path that you had no valid license at all, and for past GPL offenders that would have been the worse outcome. AI companies could change that
> But open weights don't undo the infringement -- they just make a potentially infringing artifact publicly available.
This is true when talking about the infringement of the copyrights of others. But when discussing the infringement of GPL copyleft, making a potentially infringing artifact publicly available likely satisfies the license conditions.
The evil is that this case was settled, and before being settled was decided in a way contrary to all previous copyright decisions. The courts decided that rap records had to clear every single sample, thereby basically destroying the art form, but now you can literally feed every book into a blender, piece another book together out of the pieces, and sell it.
Hip-hop when it peaked with the Bomb Squad was such a frenetic mix of so many recognizable, unrecognizable, and transformed sources that it doesn't resemble anything that was made after the decisions against Biz Markie and De La Soul. Afterwards, you just licensed one song, slightly cut it up, and rapped over it. It was just a new way to sell old shit to young people unfamiliar with it.
Now you can literally just train a machine on the same stuff, and it's legal. A machine transformation was elevated over human creativity, simply because rich people wanted it.
> The framing of 'share your weights freely' as a remedy is interesting but underspecified. The FSF's argument is essentially that training on copyrighted code without permission is infringement, and the remedy should be open weights.
Ignoring the fact that the statement doesn't talk about FSF code in the training data at all, [0] are you sure about that? From the start of the last of three paragraph in the statement:
Obviously, the right thing to do is protect computing freedom: share complete training inputs with every user of the LLM, together with the complete model, training configuration settings, and the accompanying software source code. Therefore, we urge Anthropic and other LLM developers that train models using huge datasets downloaded from the Internet to provide these LLMs to their users in freedom.
This seems to me to be consistent with the FSF's stance of "You told the computer how to do it. The right thing to do is to give the humans operating that computer the software, input data, and instructions that they need to do it, too.".
[0] In fact, it talks about the inclusion of a book published under the terms of the GNU FDL, [1] which requires distribution of modified copies of a covered work to -themselves- be covered by the GNU FDL.
It looks like the stance of FSF is for proliferation of the copyleft to trained LLMs
> "Therefore, we urge Anthropic and other LLM developers that train models using huge datasets downloaded from the Internet to provide these LLMs to their users in freedom"
No, it looks like the stance of the FSF is that models should be free as a matter of principle, the same as their stance when it comes to software. Nothing in the linked post contradicts the description that the judgement was that the training was fair use.
What weak, counter-productive, messaging. This is like having a bully punching you in the face and responding with “hey man, I’m not going to do anything about this, I’m not even going to tell an adult, but I’d urge you to consider not punching me in the face”. Great news for the bully! You just removed one concern from their mind, essentially giving the permission to be as bad to you as they want.
They could not mention they usually don’t sue and that they are small and “have to pick [their] battles”, which effectively means “there will be no repercussions from our side, we won’t even consider trying, so continue to do as you please and even worse”.
Saying nothing is an option. It is very possible (and the FSF has done it) to put yourself into a weaker position by saying something.
You don’t have to lie, but you don’t have to unpromptedly volunteer you don’t have a hand to play, either.
A related topic that I have in the past thought about is, whether LLM derived code would necessitate the release under a copyleft license because of the training data. Never saw a cogent analysis that explained either why or why not this is the case beyond practicality due to models having been utilized in closed source codebases already…
The short answer is that we don't know. The longer answer based purely on this case is that there's an argument that training is fair use and so copyleft doesn't have any impact on the model, but this is one case in California and doesn't inherently set precedent in the US in general and has no impact at all on legal interpretations in other countries.
The dearth of case law here still makes a negative outcome for FSF pretty dangerous, even if they don't appeal it and set precedent in higher courts. It might not be binding but every subsequent case will be able to site it, potentially even in other common law countries that lack case law on the topic.
And then there is the chilling effect. If FSF can't enforce their license, who is going to sue to overturn the precedent? Large companies, publishers, and governments have mostly all done deals with the devil now. Joe Blow random developer is going to get a strip mall lawyer and overturn this? Seems unlikely
I don't think this argument is a winner. It fails on a few grounds:
First, unless you can point to regurgitation of memorized code, you're not able to make an argument about distribution or replication. This is part of the problem that most publishers are having with prose text and LLMs. Modern LLMs don't memorize harry potter like GPT3 did. The memorization older models showed came from problems in the training data, e.g. harry potter and people writing about harry potter are extraordinarily over-represented. It's similar to how with stable diffusion you could prompt for anything in the region of "Van Gogh's Starry Night" and get it, since it was in the training data 50-100 different ways. You can't reliably do this with Opus or GPT5. If they're not redistributing the code verbatim, they're not in violation of the license. One could argue that the models produce "derivative works, but..."
The derivative works argument is inapt. The point of it is to disrupt someone's end-run around the license by saying that building on top of GPL code is not enough to non-GPL it. We imagine this will still work for LLMs because of the GPLs virality--I can't enclose a critical GPL module in non-GPL code and not release the GPL code. But the models aren't DOING THAT. They're not reaching for XYZ GPL'd project to build with. They're vibing out a sparsely connected network of information about literally trillions of lines of software. What comes out is a mishmash of code from here and there, and only coincidentally resembles GPL code, when it does. In order to make this argument work, you need a theory of how LLMs are trained and operate that supports it. Regardless of whether or not one of those theories exist, in court, you'd need to show that your theory was better than the company's expert witness's theory. Good luck.
Second, infringement would need discovery to uncover and would be contingent on user input. This is why the NYT sued for deleted user prompts to ChatGPT--the plaintiffs can't show in public that the content is infringing, so they need to seek discovery to find evidence. That's only going to work in cases where you survive a motion to dismiss--which is EXACTLY where a few of these suits have failed. You need to show first that you can succeed on the merits, then you proceed. That will cut down many of these challenges since they just can't show the actual infringement.
Third, and I think this is the most important, the license protections here are enforced by *copyright*. For copyright it very much matters if something is lifted verbatim vs modified. It is not like patent protection where things like clean room design are shown to have mattered to real courts on real matters. In additional contrast to patents, copyright doesn't care if the outcome is close. That's very much a concern for patents. If I patent a gizmo and you produce a gizmo that operates through nearly identical mechanisms to those I patented, then you can be sued--they don't need to be exact. If I write a novel about a boy wizard with glasses who takes a train to a school in Scotland and you write a novel about a boy wizard with glasses who takes a boat to a school in Inishmurray, I can't sue you for copyright infringement. You need to copy the words I wrote and distribute them to rise to a violation.
> Modern LLMs don't memorize harry potter like GPT3 did. [...] You can't reliably do this with Opus or GPT5.
If you try any modern LLM, you will find that you can. Easily [0], reliably [1], consistently [2]. All these examples are with models released in 2025/26.
> unless you can point to regurgitation of memorized code
I have, on many occasions, gotten an LLM to do just this. It's not particularly hard. In the most recent case google's search bar LLM happily regurgitated a digital ocean article as if it was it's own output. Searching for some strings in the comments located the original page and it was a 95% match between origin and output.
> The memorization older models showed came from problems in the training data,
And what proof do you have that they "fixed" this? And what was the fix?
> harry potter and people writing about harry potter
I'm not sure that's how you get GPT to reproduce upwards of 85% of Harry Potter novels.
> Second, infringement would need discovery to uncover and would be contingent on user input.
That's not at all how copyright infringement works. That would be if you wanted to prove malice and get triple damages. Copyright infringement is an exceptionally simple violation of the law. You either copied, or you did not.
> For copyright it very much matters if something is lifted verbatim vs modified.
Transformation is a valid defense for _some_ uses. It is not for commercial uses. Using LLM generated code for commercial purposes is a hazard.
Huh, I've been waiting for the FSF to say something about the current big issue: mandatory Operating System age-asking. Maybe now that they've meddled in a copyright lawsuit that has no broader ramifications for the public (the people they supposedly fight for), they can get back to that.
How dare they? Defending freedom of these filthy people and dignity of authors against these nice familiar corporations!
The rephrased¹ title "FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Free" certainly doesn’t dramatise enough how odious an act it can be.
¹ Original title is "The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom"
> Among the works we hold copyrights over is Sam Williams and Richard Stallman's Free as in freedom: Richard Stallman's crusade for free software, which was found in datasets used by Anthropic as training inputs for their LLMs.
This is the reason why AI companies won't let anyone inspect which content was in the training set. It turns out the suspicions from many copyright holders (including the FSF) was true (of course).
Anthropic and others will never admit it, hence why they wanted to settle and not risk going to trial. AI boosters obviously will continue to gaslight copyright holders to believe nonsense like: "It only scraped the links, so AI didn't directly train on your content!", or "AI can't see like humans, it only see numbers, binary or digits" or "AI didn't reproduce exactly 100% of the content just like humans do when tracing from memory!".
They will not share the data-set used to train Claude, even if it was trained on AGPLv3 code.
There's already legal requirements in the EU that you must publish what goes into your training set. This information must apparently be publshed before the august 2 next year.
They simply have way too much incentive to train on anything they can get their hands on. They are driving businesses, that are billions in losses so far. Someone somewhere is probably being told to feed the monster anything they can get, and not to document it, threatened with an NDA and personal financial ruin, if the proof of it ever came out. Opaque processes acting as a shield, like they do in so many other businesses.
Good. I want to see more lawsuits going after these hyper scalers for blatantly disregarding copyright law while simultaneously benefiting from it. In a just world they would all go down and we would be left with just the OSS models. But we don't live in a fair world :(
Where's the threat? The FSF was notified that as part of the settlement in Bartz v. Anthropic they were potentially entitled to money, but in this case the works in question were released under a license that allowed free duplication and distribution so no harm was caused. There's then a note that if the FSF had been involved in such a suit they'd insist on any settlement requiring that the trained model be released under a free license. But they weren't, and they're not.
(Edit: In the event of it being changed to match the actual article title, the current subject line for this thread is " FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Freel")
> but in this case the works in question were released under a license that allowed free duplication and distribution so no harm was caused.
FSF licenses contain attribution and copyleft clauses. It's "do whatever you want with it provided that you X, Y and Z". Just taking the first part without the second part is a breach of the license.
It's like renting a car without paying and then claiming "well you said I can drive around with it for the rest of the day, so where is the harm?" while conveniently ignoring the payment clause.
You maybe confusing this with a "public domain" license.
If what you do with a copyrighted work is covered by fair use it doesn't matter what the license says - you can do it anyway. The GFDL imposes restrictions on distribution, not copying, so merely downloading a copy imposes no obligation on you and so isn't a copyright infringement either.
I used to be on the FSF board of directors. I have provided legal testimony regarding copyleft licenses. I am excruciatingly aware of the difference between a copyleft license and the public domain.
> I am excruciatingly aware of the difference between a copyleft license and the public domain.
Then why did you say "no harm was caused"? Clearly the harm of "using our copylefted work to create proprietary software" was caused. Do you just mean economic harm? If so, I think that's where the parent comments confusion originates.
> The GFDL imposes restrictions on distribution, not copying, so merely downloading a copy imposes no obligation on you and so isn't a copyright infringement either.
The restrictions fall not only on verbatim distribution, but derivative works too. I am not aware whether model outputs are settled to be or not to be (hehe) derivative works in a court of law, but that question is at the vey least very much valid.
> the district court ruled that using the books to train LLMs was fair use but left for trial the question of whether downloading them for this purpose was legal.
The pipeline is something like: download material -> store material -> train models on material -> store models trained on material -> serve output generated from models.
These questions focus on the inputs to the model training, the question I have raised focuses on the outputs of the model. If [certain] outputs are considered derivative works of input material, then we have a cascade of questions which parts of the pipeline are covered by the license requirements. Even if any of the upstream parts of this simplified pipeline are considered legal, it does not imply that that the rest of the pipeline is compliant.
Consider the net effect and the answer is clear. When these models are properly "trained", are people going to look for the book or a derivative of it, with proper attribution?
Or is the LLM going to regurgitate the same content with zero attribution, and shift all the traffic away from the original work?
When viewed in this frame, it is obvious that the work is derivative and then some.
That is your opinion, but the judge disagreed with you. The decision may have been overturned on appeal, but as it stands, in that courtroom, the training was fair use.
I can memorize a song and it will be fair use too, but it won't be anymore once I start performing it publicly. Training itself is quite obviously fair use, what matters is what happens next.
This is also, unfortunately, the only way this can be settled. Making LLM output legally a derivative work would murder the AI golden rush and nobody wants that
I'm also skeptical that it's impossible to get an LLM to reproduce some code verbatim. Google had that paper a while back about getting diffusion models to spit out images that were essentially raw training data, and I wouldn't be surprised if the same is possible for LLMs.
Stack Overflow has verbatim copied GPL code in some of its questions and answers. As presented by SO, that code is not under the GPL license (this also applies to other licenses - the BSD advertising clause and the original json will cause similar problems).
Arguably, the use of the code in the Stack Overflow question and answer is fair use.
The problem occurs not when someone reads the Q&A with the improperly licensed code but rather when they then copy that code verbatim into their own non GPL product and distribute that without adherence to the GPL.
It's the last step - some human distributing the improperly licensed software that is the violation of the GPL.
This same chain of what is allowed and what is not is equally applicable to LLMs. Providing examples from GPL licensed material to answer a question isn't a license violation. The human copying that code (from any source) and pasting it into their own software is a license violation.
---
Some while back I had a discussion with a Swiss developer about the indefinite article used before "hobbit" in a text game. They used "an hobbit" and in the discussion of fixing it, I quoted the first line of The Hobbit. "In a hole in the ground there lived a hobbit." That cleared it up and my use of it in that (and this) discussion is fair use.
If someone listening to that conversation (or reading this one) thought that the bit that I quoted would be great on a T-shirt and them printed that up and distributed it - that would be a copyright violation.
The Ninth Circuit did, however, overturn the district court's decision that Google's thumbnail images were unauthorized and infringing copies of Perfect 10's original images. Google claimed that these images constituted fair use, and the circuit court agreed. This was because they were "highly transformative."
If I was to then take those thumbnails from a google image search and distribute that as an icon library, I would then be guilty of copyright infringement.
I believe that Stack Overflow, Google Images, and LLM models and their output constitutes an example of transformative fair use. What someone does with that output is where copyright infringement happens.
My claim isn't that AI vendors are blameless but rather that in the issue of copyright and license adherence it is the human in the process that is the one who has agency and needs to follow copyright (and for AI agents that were unleashed without oversight, it is the human that spun them up or unleashed them).
That's really interesting. I'm a lawyer, and I had always interpreted the license like a ToS between the developers. That (in my mind) meant that the license could impose arbitrary limitations above the default common law and statutory rules and that once you touched the code you were pregnant with those limitations, but this does make sense. TIL. So, thanks.
Does the reasoning in the cases where people to whom GPL software was distributed could sue the distributor for source code, rather than relying on the copyright holder suing for breach of copyright strengthen the argument that arbitrary limitations are enforceable?
Unrelated question regarding this part, since you seem to be an expert on this:
> If what you do with a copyrighted work is covered by fair use it doesn't matter what the license says - you can do it anyway.
How is it that contracts can prohibit trial by jury but they can't ban prohibit fair use of copyrighted work? Is there a list of things a contract is and isn't allows to prohibit, and explanations/reasons for them?
The general answer is because there is a statute or court opinion that says so for one thing and a different one that says something else for the other thing.
It's also relevant that copyright (and fair use) is federal law, contracts are state law and federal law preempts state law.
It sounds that way a bit from the one sentence. But that’s not the case at all.
> 4. MODIFICATIONS
> You may copy and distribute a Modified Version of the Document under the conditions of sections 2 and 3 above, provided that you release
the Modified Version under precisely this License, with the Modified
Version filling the role of the Document, thus licensing distribution
and modification of the Modified Version to whoever possesses a copy
of it. In addition, you must do these things in the Modified Version:
Etc etc.
In short, it is a copyleft license. You must also license derivative works under this license.
Just fyi, the gnu fdl is (unsurprisingly) available for free online - so if you want to know what it says, you can read it!
And the judgement said that the training was fair use, but that the duplication might be an infringement. The GFDL doesn't restrict duplication, only distribution, so if training on GFDLed material is fair use and not the creation of a derivative work then there's no damage.
Right. I can publish the work in whole without asking permission. That’s unrestricted duplication.
However, as i read it, an LLM spitting out snippets from the text is not “duplicating” the work. That would fall under modifications. From the license:
> A "Modified Version" of the Document means any work containing the Document or a portion of it, either copied verbatim, or with modifications and/or translated into another language.
I read that pretty clearly as any work containing text from a gnu fdl document is a modification not a duplication.
1) Obtaining the copyrighted works used for training. Anthropic did this without asking for the copyright holders' permission, which would be a copyright violation for any work that isn't under a license that grants permission to duplicate. The GFDL does, so no issue here.
2) Training the model. The case held that this was fair use, so no issue here.
3) Whether the output is a derivative work. If so then you get to figure out how the GFDL applies to the output, but to the best of my knowledge the case didn't ask this question so we don't know.
For this to stand up in court you'd need to show that an LLM is distributing "a modified version of the document".
If I took a book and cut it up into individual words (or partial words even), and then used some of the words with words from every other book to write a new book, it'd be hard to argue that I'm really "distributing the first book", even if the subject of my book is the same as the first one.
This really just highlights how the law is a long way behind what's achievable with modern computing power.
You’re just describing transformative use. I’m not a lawyer, but an example from music - taking a single drum hit from a james brown song is apparently not transformative. Taking a vibe from another song is also maybe not transformative, e.g. robin thicke and pharrell’s “blurred lines” was found to legally take the “feel” from Marvin Gaye’s “Got to Give it Up”
Which is all to say that the law is actually really bad at determining what is right and wrong, and our moral compasses should not defer to the law. Unfortunately, moral compasses are often skewed by money - like how normal compassess are skewed by magnets
Presumably, a suitable prompt could get the LLM to produce whole sections of the book which would demonstrate that the LLM contains a modified version.
I am distrubting an svg file. It’s a program that, when run, produces an image of mickey mouse.
By your description of the law, this svg file is not infringing on disney’s copyright - since it’s a program that when run creates an infringing document (the rasterized pixels of mickey mouse) but it is not an infringing document itself.
I really don’t think my “i wrote a program in the svg language” defense would hold up in court. But i wonder how many levels of abstraction before it’s legal? Like if i write the mickey-mouse-generator in python does that make it legal? If it generates a variety of randomized images of mickey mouse, is that legal? If it uses statistical anaylsis of many drawings of mickey to generate an average mickey mouse, is that legal? Does it have to generate different characters if asked before it is legal? Can that be an if statement or does it have to use statistical calculations to decide what character i want?
I don't like the editorialized title either but I would say that the actual post title
"The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom"
and this sentence at the end
" We are a small organization with limited resources and we have to pick our battles, but if the FSF were to participate in a lawsuit such as Bartz v. Anthropic and find our copyright and license violated, we would certainly request user freedom as compensation."
Is it? The FSF's description of the judgement is that the training was fair use, but that the actual downloading of the material may have been a copyright infringement. What software does the FSF hold copyright to that can't be downloaded freely? Under what circumstances would the FSF be in a position to influence the nature of a settlement if they weren't harmed?
Copyright infringement causes harm, so if there's no harm there's no infringement. You can freely duplicate GFDLed material, so downloading it isn't an infringement. If training a model on that downloaded material is fair use then there's no infringement.
Is the FSF threatening Anthropic? The way I read it looks like they are not:
> We are a small organization with limited resources and we have to pick our battles, but if the FSF were to participate in a lawsuit such as Bartz v. Anthropic and find our copyright and license violated, we would certainly request user freedom as compensation.
Sounds more like “we can’t and won’t sue, but this is the kind of compensation that we think would be appropriate”
The rule is fine and clear, it just wasn’t followed here. There’s no reason to have a stricter rule, what you’re complaining about is its enforcement. Two moderators can’t read everything, if you have a complaint, email them (contact link at the bottom of the page), they are quite responsive.
The hero we need, but not the hero we deserve..
The issue is that every CS masters student & AI researcher knows how to build a SOTA LLM.. But, only a few companies have the resources.
The process:
(1) steal as much data from the internet as possible (data is everything) (2) raise incomprehensible amounts of money (3) find a location where you can take over the energy grid for training (4) put a black box around it so nobody can see the weights (5) charge users $$$ to use (6) retrain models with user session data (opt in by default) (7) peek around at how users are using, (maybe) change policies to stop them from using that way, and (maybe) rapidly develop features for that use case.
(Sorry that last one is jaded and not fair - just included to give you a picture of what could be happening with this sort of tech) …
The entire premise of the product is “built on the backs of any & everyone who has ever published a work”
Do any products exist which are not built on uncompensated work of other people in the past?
Generally speaking societies do better when knowledge is shared and not hoarded.
Hoarding knowledge via legal constructs is great at concentrating wealth to the hoarder at the expense of everyone else.
We should restore copyright to its original term lengths.
I agree with the stance of Anthropic et al that these models should be built with all possible information.
I agree with the stance of the FSF that the resulting models should be as freely usable/available as possible.
These companies do even better because we're not allowed to share the knowledge (read, illegally copy protected works) and they are.
It would be nice if members of the class could vote to force a case to trial. For the typical token settlement amount, I’m sure many would rather have the precedent-setting case instead.
But whether you can actually be compelled to do that isn't well tested in court. Challenging that the GPL is enforcable in that way leads you down the path that you had no valid license at all, and for past GPL offenders that would have been the worse outcome. AI companies could change that
This is true when talking about the infringement of the copyrights of others. But when discussing the infringement of GPL copyleft, making a potentially infringing artifact publicly available likely satisfies the license conditions.
The evil is that this case was settled, and before being settled was decided in a way contrary to all previous copyright decisions. The courts decided that rap records had to clear every single sample, thereby basically destroying the art form, but now you can literally feed every book into a blender, piece another book together out of the pieces, and sell it.
Hip-hop when it peaked with the Bomb Squad was such a frenetic mix of so many recognizable, unrecognizable, and transformed sources that it doesn't resemble anything that was made after the decisions against Biz Markie and De La Soul. Afterwards, you just licensed one song, slightly cut it up, and rapped over it. It was just a new way to sell old shit to young people unfamiliar with it.
Now you can literally just train a machine on the same stuff, and it's legal. A machine transformation was elevated over human creativity, simply because rich people wanted it.
Ignoring the fact that the statement doesn't talk about FSF code in the training data at all, [0] are you sure about that? From the start of the last of three paragraph in the statement:
This seems to me to be consistent with the FSF's stance of "You told the computer how to do it. The right thing to do is to give the humans operating that computer the software, input data, and instructions that they need to do it, too.".[0] In fact, it talks about the inclusion of a book published under the terms of the GNU FDL, [1] which requires distribution of modified copies of a covered work to -themselves- be covered by the GNU FDL.
[1] <https://www.gnu.org/licenses/fdl-1.3.html>
> "Therefore, we urge Anthropic and other LLM developers that train models using huge datasets downloaded from the Internet to provide these LLMs to their users in freedom"
Saying nothing is an option. It is very possible (and the FSF has done it) to put yourself into a weaker position by saying something.
You don’t have to lie, but you don’t have to unpromptedly volunteer you don’t have a hand to play, either.
And then there is the chilling effect. If FSF can't enforce their license, who is going to sue to overturn the precedent? Large companies, publishers, and governments have mostly all done deals with the devil now. Joe Blow random developer is going to get a strip mall lawyer and overturn this? Seems unlikely
First, unless you can point to regurgitation of memorized code, you're not able to make an argument about distribution or replication. This is part of the problem that most publishers are having with prose text and LLMs. Modern LLMs don't memorize harry potter like GPT3 did. The memorization older models showed came from problems in the training data, e.g. harry potter and people writing about harry potter are extraordinarily over-represented. It's similar to how with stable diffusion you could prompt for anything in the region of "Van Gogh's Starry Night" and get it, since it was in the training data 50-100 different ways. You can't reliably do this with Opus or GPT5. If they're not redistributing the code verbatim, they're not in violation of the license. One could argue that the models produce "derivative works, but..."
The derivative works argument is inapt. The point of it is to disrupt someone's end-run around the license by saying that building on top of GPL code is not enough to non-GPL it. We imagine this will still work for LLMs because of the GPLs virality--I can't enclose a critical GPL module in non-GPL code and not release the GPL code. But the models aren't DOING THAT. They're not reaching for XYZ GPL'd project to build with. They're vibing out a sparsely connected network of information about literally trillions of lines of software. What comes out is a mishmash of code from here and there, and only coincidentally resembles GPL code, when it does. In order to make this argument work, you need a theory of how LLMs are trained and operate that supports it. Regardless of whether or not one of those theories exist, in court, you'd need to show that your theory was better than the company's expert witness's theory. Good luck.
Second, infringement would need discovery to uncover and would be contingent on user input. This is why the NYT sued for deleted user prompts to ChatGPT--the plaintiffs can't show in public that the content is infringing, so they need to seek discovery to find evidence. That's only going to work in cases where you survive a motion to dismiss--which is EXACTLY where a few of these suits have failed. You need to show first that you can succeed on the merits, then you proceed. That will cut down many of these challenges since they just can't show the actual infringement.
Third, and I think this is the most important, the license protections here are enforced by *copyright*. For copyright it very much matters if something is lifted verbatim vs modified. It is not like patent protection where things like clean room design are shown to have mattered to real courts on real matters. In additional contrast to patents, copyright doesn't care if the outcome is close. That's very much a concern for patents. If I patent a gizmo and you produce a gizmo that operates through nearly identical mechanisms to those I patented, then you can be sued--they don't need to be exact. If I write a novel about a boy wizard with glasses who takes a train to a school in Scotland and you write a novel about a boy wizard with glasses who takes a boat to a school in Inishmurray, I can't sue you for copyright infringement. You need to copy the words I wrote and distribute them to rise to a violation.
If you try any modern LLM, you will find that you can. Easily [0], reliably [1], consistently [2]. All these examples are with models released in 2025/26.
[0] https://arxiv.org/html/2601.02671?amp=&=
[1] https://arxiv.org/abs/2506.12286
[2] https://ai.stanford.edu/blog/verbatim-memorization/
I have, on many occasions, gotten an LLM to do just this. It's not particularly hard. In the most recent case google's search bar LLM happily regurgitated a digital ocean article as if it was it's own output. Searching for some strings in the comments located the original page and it was a 95% match between origin and output.
> The memorization older models showed came from problems in the training data,
And what proof do you have that they "fixed" this? And what was the fix?
> harry potter and people writing about harry potter
I'm not sure that's how you get GPT to reproduce upwards of 85% of Harry Potter novels.
> Second, infringement would need discovery to uncover and would be contingent on user input.
That's not at all how copyright infringement works. That would be if you wanted to prove malice and get triple damages. Copyright infringement is an exceptionally simple violation of the law. You either copied, or you did not.
> For copyright it very much matters if something is lifted verbatim vs modified.
Transformation is a valid defense for _some_ uses. It is not for commercial uses. Using LLM generated code for commercial purposes is a hazard.
They don't have the rights to distribute the training data.
The rephrased¹ title "FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Free" certainly doesn’t dramatise enough how odious an act it can be.
¹ Original title is "The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom"
This is the reason why AI companies won't let anyone inspect which content was in the training set. It turns out the suspicions from many copyright holders (including the FSF) was true (of course).
Anthropic and others will never admit it, hence why they wanted to settle and not risk going to trial. AI boosters obviously will continue to gaslight copyright holders to believe nonsense like: "It only scraped the links, so AI didn't directly train on your content!", or "AI can't see like humans, it only see numbers, binary or digits" or "AI didn't reproduce exactly 100% of the content just like humans do when tracing from memory!".
They will not share the data-set used to train Claude, even if it was trained on AGPLv3 code.
(Edit: In the event of it being changed to match the actual article title, the current subject line for this thread is " FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Freel")
FSF licenses contain attribution and copyleft clauses. It's "do whatever you want with it provided that you X, Y and Z". Just taking the first part without the second part is a breach of the license.
It's like renting a car without paying and then claiming "well you said I can drive around with it for the rest of the day, so where is the harm?" while conveniently ignoring the payment clause.
You maybe confusing this with a "public domain" license.
I used to be on the FSF board of directors. I have provided legal testimony regarding copyleft licenses. I am excruciatingly aware of the difference between a copyleft license and the public domain.
Then why did you say "no harm was caused"? Clearly the harm of "using our copylefted work to create proprietary software" was caused. Do you just mean economic harm? If so, I think that's where the parent comments confusion originates.
The restrictions fall not only on verbatim distribution, but derivative works too. I am not aware whether model outputs are settled to be or not to be (hehe) derivative works in a court of law, but that question is at the vey least very much valid.
> the district court ruled that using the books to train LLMs was fair use but left for trial the question of whether downloading them for this purpose was legal.
The pipeline is something like: download material -> store material -> train models on material -> store models trained on material -> serve output generated from models.
These questions focus on the inputs to the model training, the question I have raised focuses on the outputs of the model. If [certain] outputs are considered derivative works of input material, then we have a cascade of questions which parts of the pipeline are covered by the license requirements. Even if any of the upstream parts of this simplified pipeline are considered legal, it does not imply that that the rest of the pipeline is compliant.
Or is the LLM going to regurgitate the same content with zero attribution, and shift all the traffic away from the original work?
When viewed in this frame, it is obvious that the work is derivative and then some.
Licences like AGPL also don't have redistribution as their only restriction.
Arguably, the use of the code in the Stack Overflow question and answer is fair use.
The problem occurs not when someone reads the Q&A with the improperly licensed code but rather when they then copy that code verbatim into their own non GPL product and distribute that without adherence to the GPL.
It's the last step - some human distributing the improperly licensed software that is the violation of the GPL.
This same chain of what is allowed and what is not is equally applicable to LLMs. Providing examples from GPL licensed material to answer a question isn't a license violation. The human copying that code (from any source) and pasting it into their own software is a license violation.
---
Some while back I had a discussion with a Swiss developer about the indefinite article used before "hobbit" in a text game. They used "an hobbit" and in the discussion of fixing it, I quoted the first line of The Hobbit. "In a hole in the ground there lived a hobbit." That cleared it up and my use of it in that (and this) discussion is fair use.
If someone listening to that conversation (or reading this one) thought that the bit that I quoted would be great on a T-shirt and them printed that up and distributed it - that would be a copyright violation.
Google's use of thumbnails for images was found to be fair use. https://en.wikipedia.org/wiki/Perfect_10,_Inc._v._Amazon.com...
If I was to then take those thumbnails from a google image search and distribute that as an icon library, I would then be guilty of copyright infringement.I believe that Stack Overflow, Google Images, and LLM models and their output constitutes an example of transformative fair use. What someone does with that output is where copyright infringement happens.
My claim isn't that AI vendors are blameless but rather that in the issue of copyright and license adherence it is the human in the process that is the one who has agency and needs to follow copyright (and for AI agents that were unleashed without oversight, it is the human that spun them up or unleashed them).
> If what you do with a copyrighted work is covered by fair use it doesn't matter what the license says - you can do it anyway.
How is it that contracts can prohibit trial by jury but they can't ban prohibit fair use of copyrighted work? Is there a list of things a contract is and isn't allows to prohibit, and explanations/reasons for them?
It's also relevant that copyright (and fair use) is federal law, contracts are state law and federal law preempts state law.
"Sam Williams and Richard Stallman's Free as in freedom: Richard Stallman's crusade for free software"
"GNU Free Documentation License (GNU FDL). This is a free license allowing use of the work for any purpose without payment."
I'm not familiar with this license or how it compares to their software licenses, but it sounds closer to a public domain license.
> 4. MODIFICATIONS
> You may copy and distribute a Modified Version of the Document under the conditions of sections 2 and 3 above, provided that you release the Modified Version under precisely this License, with the Modified Version filling the role of the Document, thus licensing distribution and modification of the Modified Version to whoever possesses a copy of it. In addition, you must do these things in the Modified Version:
Etc etc.
In short, it is a copyleft license. You must also license derivative works under this license.
Just fyi, the gnu fdl is (unsurprisingly) available for free online - so if you want to know what it says, you can read it!
Right. I can publish the work in whole without asking permission. That’s unrestricted duplication.
However, as i read it, an LLM spitting out snippets from the text is not “duplicating” the work. That would fall under modifications. From the license:
> A "Modified Version" of the Document means any work containing the Document or a portion of it, either copied verbatim, or with modifications and/or translated into another language.
I read that pretty clearly as any work containing text from a gnu fdl document is a modification not a duplication.
1) Obtaining the copyrighted works used for training. Anthropic did this without asking for the copyright holders' permission, which would be a copyright violation for any work that isn't under a license that grants permission to duplicate. The GFDL does, so no issue here. 2) Training the model. The case held that this was fair use, so no issue here. 3) Whether the output is a derivative work. If so then you get to figure out how the GFDL applies to the output, but to the best of my knowledge the case didn't ask this question so we don't know.
If I took a book and cut it up into individual words (or partial words even), and then used some of the words with words from every other book to write a new book, it'd be hard to argue that I'm really "distributing the first book", even if the subject of my book is the same as the first one.
This really just highlights how the law is a long way behind what's achievable with modern computing power.
Which is all to say that the law is actually really bad at determining what is right and wrong, and our moral compasses should not defer to the law. Unfortunately, moral compasses are often skewed by money - like how normal compassess are skewed by magnets
By your description of the law, this svg file is not infringing on disney’s copyright - since it’s a program that when run creates an infringing document (the rasterized pixels of mickey mouse) but it is not an infringing document itself.
I really don’t think my “i wrote a program in the svg language” defense would hold up in court. But i wonder how many levels of abstraction before it’s legal? Like if i write the mickey-mouse-generator in python does that make it legal? If it generates a variety of randomized images of mickey mouse, is that legal? If it uses statistical anaylsis of many drawings of mickey to generate an average mickey mouse, is that legal? Does it have to generate different characters if asked before it is legal? Can that be an if statement or does it have to use statistical calculations to decide what character i want?
wikipedia used to be under FDL and they lobbied FSF to allow an escape hatch to Commons for a few months, because FDL was so annoying.
"The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom"
and this sentence at the end
" We are a small organization with limited resources and we have to pick our battles, but if the FSF were to participate in a lawsuit such as Bartz v. Anthropic and find our copyright and license violated, we would certainly request user freedom as compensation."
could be seen as "threatening".
Not a nothing burger, but not totally insignificant either.
> We are a small organization with limited resources and we have to pick our battles, but if the FSF were to participate in a lawsuit such as Bartz v. Anthropic and find our copyright and license violated, we would certainly request user freedom as compensation.
Sounds more like “we can’t and won’t sue, but this is the kind of compensation that we think would be appropriate”
The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom