> If source code can now be generated from a specification, the specification is where the essential intellectual content of a GPL project resides.
Our foreparents fought for the right to implement works-a-like to corporate software packages, even if the so-called owners did not like it. We're ready to throw it all away, and let intellectual property owners get so much more control.
The implications will not end up being anti-large-corporation or pro-sharing. If you can prevent someone from re-implementing a spec or building a client that speaks your API or building a work-a-like, it will be the large corporations that exersize this power as usual.
And the whole Adobe pdf thing and the whole Microsoft word thing. And the whole ibm pc thing. Imagine if we were forced to keep using ibm from when they lost their way until now simply because anti-ai luddites were able to scare monger
I would wager that the vast majority of people commenting here about the pitfalls of AI, especially as it relates to governance and laws, are heavy users of AI, recognize the import and value it brings, and find ways to utilize it more themselves, so not sure using an ad-hominem dismissal of very valid objections are going to be effective.
(side bar: the phrase "anti-<whatever> luddites" is way, way overused, especially here. Let's get more creative, people!)
>Our foreparents fought for the right to implement works-a-like to corporate software packages, even if the so-called owners did not like it
Our "foreparents" weren't competing with corporations with unlimited access to generative AI trained on their work. The times, they're-a-changin'.
You're rehashing the argument made in one of the articles which this piece criticizes and directly addresses, while ignoring the entirety of what was written before the conclusion that you quoted.
If anyone finds themselves agreeing with the comment I'm responding to, please, do yourself a favor and read the linked article.
I would do no justice to it by reiterating its points here.
I believe the GP post is saying that if we react to the new AI-enabled environment by arbitrarily strengthening IP controls for IP owners, the greatest benefactors will almost certainly be lawyer-laden corporations, not communities, artists, or open source projects. That seems like a reasonable argument.
It seems like the answer is to adjust IP owner rights very carefully, if that's possible. It sounds very hard, though.
I mean. Yeah. GPL's genius was that it used Copyright, which proprietary enterprise wouldn't dare dismantle, to secure for the public a permanent public good.
Pretty sure no one, (but me anyway) saw overt theft of IP by ignoring IP law through redefinition coming. Admittedly I couldn't articulate for you capital would skill transfer and commoditize it in the form of pay to play data centers, but give me a break, I was a teenager/twenty something at the time.
Would software be more or less free in a world without copyright?
I argue more free. EULAs and restrictions on how+for what software can be used, like DRM, typically use copyright as their legal backing. GPL licenses turn that on it's head but that doesn't redeem the original, flawed, law.
This seems to follow the letter but not the spirit of the license. If this does pass legal muster, we can do the same to whatever proprietary software we wish, which makes a dramatically different but IMO better ecosystem in the end.
The really interesting question to me is if this transcends copyright and unravels the whole concept of intellectual property. Because all of it is premised on an assumption that creativity is "hard". But LLMs are not just writing software, they are rapidly being engineered to operate completely generally as knowledge creation engines: solving math proofs, designing drugs, etc.
So: once it's not "hard" any more, does IP even make sense at all? Why grant monopoly rights to something that required little to no investment in the first place? Even with vestigial IP law - let's say, patents: it just becomes and input parameter that the AI needs to work around the patents like any other constraints.
> So: once it's not "hard" any more, does IP even make sense at all? Why grant monopoly rights to something that required little to no investment in the first place? Even with vestigial IP law - let's say, patents: it just becomes and input parameter that the AI needs to work around the patents like any other constraints.
I think it still does: IIRC, the current legal situation is AI-output does not qualify for IP protections (at least not without substantial later human modification). IP protections are solely reserved for human work.
And I'm fine with that: if a person put in the work, they should have protections so their stuff can't be ripped off for free by all the wealthy major corporations that find some use for it. Otherwise: who cares about the LLMs.
I think you have a rather idealized model of IP in mind. In practice, IP law tends to be an expensive weapon the wealthy major corporations use against the little guy. Deep enough pockets and a big enough warchest of broad parents will drain the little guy every time.
> In practice, IP law tends to be an expensive weapon the wealthy major corporations use against the little guy. Deep enough pockets and a big enough warchest of broad parents will drain the little guy every time.
Then fix that instead of blowing it up. Because IP law is also literally the only thing that protects the little guy's work in many cases.
Arguments like yours are kinda unfathomably incomplete to me, almost like they're the remnants of some propaganda campaign. It's constructed to appeal to the defense of the little guy, but the actual effect would be to disempower him and further empower the wealthy major corporations with "big enough warchest[s]."
I mean, one thing I think the RIAA would love is to stop paying royalties to every artist ever. And the only thing they'd be worried about is an even bigger fish (like Amazon, Apple, or Spotify) no longer paying royalties to them. But as you said, they have a big enough war chest that they probably could force a deal somehow. All the artists without a war chest? Left out in the cold.
Blowing up IP would sink the RIAA. They would no longer have legal grounds to go after file sharing, and I’m confident that given the same legal footing that file sharing would win any day of the week.
I beg to differ. AI-output did not entitle the person creating the prompt for IP protections, so far – but my objection is not directed towards the "so far", but towards your omission of "the person creating the prompt", because if an AI outputs copyrighted material from the training data, that material is still copyrighted. AI is not a magical copyright removal machine.
What this means in practice is that (currently), all output of an LLM is legally considered to not be copyrightable (to the extent that it's an original work). If it happens to regurgitate an existing copyrighted work, though, is that infringement? I'm not sure we have a legal precedent on that question yet.
There’s several large settlements that say Anthropomorphic/OAI didn’t want to have legal precedent. In general if it’s not outright regurgitated it would be derivative.
Does this matter in practice though? By modifying some of the generated code and not taking a solution produced by an LLM end-to-end but borrowing heavily from it, can't a human claim full ownership of the IP even though in reality the LLM did most of the relevant work?
Good. Intellectual property is now a twisted concept by the elite, whatever its benefits were previously. As soon as Disney made Mickey popular, it was all downhill.
More likely: this is a transitional phase where our previously hard problems become easy, and we will soon set our sights on new and much harder problems. The pinnacle of creative achievement in the universe is probably not 2010s B2B SaaS.
It is entirely possible, however, that human beings will not be the primary drivers of progress on those problems.
It might unravel intellectual property, just not in a fair way. When capitalism started, public land was enclosed to create private property. Despite this being in many cases a quite unfair process, we still respect this arrangement.
With AI, a similar process is happening - publicly available information becomes enclosed by the model owners. We will probably get a "vestigial" intellectual property in the form of model ownership, and everyone will pay a rent to use it. In fact, companies might start to gatekeep all the information to only their own LLM flavor, which you will be required to use to get to the information. For example, product documentation and datasheets will be only available by talking to their AI.
Even if all I have to do is tell my agent, "here is a patent for a drug, analyse the patent and determine an equivalent but non-infringing drug" and it chugs away for a couple of hours and spits out a drug along with all the specifications to manufacture it?
I guess the state of play will be that for new drugs the original manufacturer will already have done that and ensured that literally anything that could be found as a workaround is included in the scope of the patent. But I feel like it will not be possible to keep that wartertight.
Yes, even so. Human drug researchers have been doing the same thing for decades. As soon as one pharmaceutical company launches a successful small-molecule drug everyone else jumps to find a minor tweak that will hit the same target (ideally with fewer side effects) while evading the patent. There is already specialized software to help with this process so I'm skeptical that LLM agents would be very helpful for this use case.
That also seems relevant for this whole discussion, actually -- if a work can't be copyrighted it certainly can't have a changed license, or any license at all. (I guess it's effectively public domain to the extent that it's public at all?)
You're really missing the point in multiple ways. First, precedents on copyright law are irrelevant to patent law. Second, AI generated works generally can be copyrighted under the human creator's name.
No, I think you are quite incorrect, at least on the latter point:
"Lower courts upheld a U.S. Copyright Office decision that the AI-crafted visual art at issue in the case was ineligible for copyright protection because it did not have a human creator."
Not eligible for copyright protection does not mean it can be copyrighted "under the human creator's name". It means there is no creative work at all. No copyright.
If you think about creative outcomes as n dimensional 'volumes', AI expressions can cover more than humans in many domains. These are precisely artistic styles, music styles etc. and tbh not everyone can be a Mozart but may be a lot more with AI can be Mozart lite. This begs the question how much of creativity is appreciated as a shared experience
I've always thought the opposite: IP law was created to make sure creativity stays hard, and hence controllable by the elites.
Patents came along when farmers started making city goods, threatening guilds secrets. Copyright came when the printing press made copying and translating the bible easy and accessible to all. (Trademark admittedly does not fit this view, but doesn't seem all that damaging either)
To Protect The Arts, and To Time Limit Trade Secrets were just the Protect The Children of old times, a way to confuse people who didn't look too hard at actual consequences.
This means that the future of IP depends on what lets the powers that be pull up the ladder behind them. Long term I'd expect e.g. copyright expansion and harder enforcement, just because cloning by AI gets easy enough to threaten the status quo.
> Trademark admittedly does not fit this view, but doesn't seem all that damaging either
Isn’t trademark the only thing keeping a certain cartoon mouse out of the public domain, despite the fact that his earliest animations are out of copyright? Not sure if you’d consider that damaging, or if anyone has yet tested the boundaries of the House of Mouse’s patience here.
Don't worry. The courts have consistently sided with huge companies on copyright. In the US. In Europe. Doesn't matter.
Company incorporates GPL code in their product? Never once have courts decided to uphold copyright. HP did that many times. Microsoft got caught doing it. And yet the GPL was never applied to their products. Every time there was an excuse. An inconsistent excuse.
Schoolkid downloads a movie? 30,000 USD per infraction PLUS armed police officer goes in and enforces removal of any movies.
Or take the very subject here. AI training WAS NOT considered fair use when OpenAI violated copyright to train. Same with Anthropic, Google, Microsoft, ... They incorporated harry potter and the linux kernel in ChatGPT, in the model itself. Undeniable. Literally. So even if you accept that it's changed now, OpenAI should still be forced to redistribute the training set, code, and everything needed to run the model for everything they did up to 2020. Needless to say ... courts refused to apply that.
So just apply "the law", right. Courts' judgement of using AI to "remove GPL"? Approved. Using AI to "make the next Disney-style movie"? SEND IN THE ARMY! Whether one or the other violates the law according to rational people? Whatever excuse to avoid that discussion is good enough.
Someone should put this to the test. Take the recently leaked Minecraft source code and have Copilot build an exact replica in another programming language and then publish it as open source. See if Microsoft believes AI is copyright infringement or not.
As described, this would not be the same thing. If the AI is looking at the source and effectively porting it, that is likely infringement. The idea instead should be "implement Minecraft from scratch" but with behavior, graphics, etc. identical. Note that you'll need to have an AI generate assets or something since you can't just reuse textures and models.
AI models have already looked at the source of GPL software and contain it in their dataset. Adding the minecraft source to the mix wouldn't seem much different. Of course art assets and trade marks would have to be replaced. But an AI "clean room" implementation has yet to be legally tested.
That's why he is saying it's not equivalent. For it to be the same, the LLM would have to train on/transform Minecraft's source code into its weights, then you prompt the LLM to make a game using the specifications of Minecraft solely through prompts. Of course it's copyright infringement if you just give a tool Minecraft's source code and tell it to copy it, just like it would be copyright infringement if you used a copier to copy Minecraft's source code into a new document and say you recreated Minecraft.
What if Copilot was already trained with Minecraft code in the dataset? Should be possible to test by telling the model to continue a snippet from the leaked code, the same way a news website proved their articles were used for training.
For copyright purposes I think there is an important legal distinction between training data (fed in once, ahead of time, and can in theory no longer be recovered as-is) and context window data (stored exactly for the duration of the model call).
I'm not sure there should be, but I think there is.
The big question is: if copyrighted material was used in the training material, is the LLM's output copyright infringement when it resembles the training material? In your example, you are taking the copyrighted material and giving it to the LLM as input and instructing the LLM to process it. Regardless of where the legal cards fall, this is a much less ambiguous scenario.
this question should've been posed earlier when first LLMs were training. many people chose to ignore the question, and now, several distillation epochs later, it is not a question that matters, as both yes/no are true, and not true.
is it legitimate for millions of people to exploit, expound on knowledge that was perhaps, to begin with, not legitimate to use? well they did already, who's to judge the commons now?
What a ridiculous take. Many people loudly raised the question and objected to the practice from the beginning, but a handful of companies ignored the objections and ran faster than the legal system. If they were in the wrong, legally or morally, they still deserve to face repercussions for it.
Very much is. "Software programs, as such" are exempt in the EPC article 52. However if the software program interacts with the world - if it has a "further technical effect" - it is patentable.
They might not care. Products win not by quality or features but by advertisement, hype and network effects.
The original implementation would still have the upper hand here. OTOH if I as a nobody create something cool, there's nothing stopping a huge corporation from "reimplementing" (=stealing) it and and using their huge advertising budget to completely overshadow me.
Given how hard companies like Nintendo and Microsoft have been taking down leaks or fan creations, it seems they very much do care about keeping this stuff locked down.
I believe it is a narrow view of the situation. If we take a look into the history, into the reasons for inventing GPL, we'll see that it was an attempt to fight copyrights with copyrights. The very name 'copyleft' is trying to convey the idea.
What AI are eroding is copyright. You can re-implement not just a GPL program, but to reverse engineer and re-implement a closed source program too, people have demonstrated it already, there were stories here on HN about it.
AI is eroding copyright, so there may no longer be a need for the GPL. GNU should stop and rethink its stance, chuck away the GPL as the main tool to fight evil software corporations and embrace LLM as the main weapon.
LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source where you could do innovative work on your own computer running Linux or FreeBSD or some other open OS.
I don't think that's an exciting idea for the Free Software Foundation.
Perhaps with time we'll be able to run local ones that are 'good enough', but we're not there yet.
There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
Edit: I guess the conclusion I come to is that LLM's are good for 'getting things done', but the context in which they are operating is one where the balance of power is heavily tilted towards capital, and open source is perhaps less interesting to participate in if the machines are just going to slurp it up and people don't have to respect the license or even acknowledge your work.
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source
Yeah, a bit of a conundrum. But I don't think that fighting for copyright now can bring any benefits for FOSS. GNU should bring Stallman back and see whether he can come with any new ideas and a new strategy. Alternatively they could try without Stallman. But the point is: they should stop and think again. Maybe they will find a way forward, maybe they won't but it means that either they could continue their fight for a freedom meaningfully, or they could just stop fighting and find some other things to do. Both options are better then fighting for copyright.
> There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
I want a clarify this statement a bit. The thing with LLM relying on work of others are not against GPU philosophy as I understand it: algorithms have to be free. Nothing wrong with training LLMs on them or on programs implementing them. Nothing wrong with using these LLMs to write new (free) programs. What is wrong are corporations reaping all the benefits now and locking down new algorithms later.
I think it is important, because copyright is deemed to be an ethical thing by many (I think for most people it is just a deduction: abiding the law is ethical, therefore copyright is ethical), but not for GNU.
IMO the primary significant trend in AI. Doesn't get talked about nearly enough. Means the AI is working, I guess.
>GNU should bring Stallman back ... Alternatively they could try without Stallman.
Leave Britney alone >:(
>copyright is deemed to be an ethical thing by many (I think for most people it is just a deduction: abiding the law is ethical, therefore copyright is ethical)
I've busted out "intellectual property is a crime against humanity" at layfolk to see if that shortcuts through that entire little politico-philosophical minefield. They emote the requisite mild shock when such things as crimes against humanity are mentioned; as well as at someone making such a radical statement which seems to come from no familiar species of echo chamber; and then a moment later they begin to very much look like they see where I'm coming from.
How do you even argue such a thing? I've had no such luck, I've met many people who seem to view copyright and a person owning their ideas and work as a sort of inherent moral.
Not saying this gets through to people, but copyright is purely about the legal ability to restrict what other people do. Whereas property rights are about not allowing others to restrict what you do (e.g. by taking your stuff).
>Perhaps with time we'll be able to run local ones that are 'good enough', but we're not there yet.
Right now, we can get local models that you can run on consumer hardware, that match capabilities of state of the art models from two years ago. The improvements to model architecture may or may not maintain the same pace in the future, but we will get a local equivalent to Opus 4.6 or whatever other benchmark of "good enough" you have, in the foreseeable future.
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones
There are near-SOTA LLM's available under permissive licenses. Even running them doesn't require prohibitive expenses on hardware unless you insist on realtime use.
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source where you could do innovative work on your own computer running Linux or FreeBSD or some other open OS.
When the FSF and GPL were created, I don't think this was really a consideration. They were perfectly happy with requiring Big Iron Unix or an esoteric Lisp Machine to use the software - they just wanted to have the ability to customize and distribute fixes and enhancements to it.
> There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
This was already the case and it just got worse, not better.
At a certain point, I think we had reached a kind of equilibrium where some corporations were decent open source citizens. They understood that they could open source things like infrastructure or libraries and keep their 'crown jewels' closed. And while Stallman types might not have been happy with that, it seemed to work out for people.
Now they've just hoovered up all the free stuff into machines that can mix it up enough to spit it out in a way that doesn't even require attribution, and you have to pay to use their machine.
AI essentially gatekeeps all of open source to companies to pluck from to their hearts content. And individual contributors using these tools and freely mixing it with their own - usual minor - contributions are another step of whitewashing because they're definitely not going to own up to writing only 5% of the stuff they got paid for.
Before we had RedHat and Ubuntu, who at least were contributing back, now we have Microsoft, Anthropic and OpenAI who are racing to lock the barn door around their new captive sheep. It's just a massive IP laundromat.
Is massive capital expenditure not also required to enforce the GPL? If some company steals your GPLed code and doesn't follow the license, you will have to sue them and somebody will have to pay the lawyers.
> Is massive capital expenditure not also required to enforce the GPL?
It's nowhere near the order of magnitude of the kind of spending they're sinking into LLM's. The FSF and other groups were reasonably successful at enforcing the GPL, operating on a budget 1000's of times smaller than that of AI companies.
Right but LLM companies are building frontier models with frontier talent while trying to sock up demand with a loss leader strategy, on top of an historic infrastructure build out.
Being able to coat efficiently run frontier models is i think, not a high priced endeavor for an org (compared to an individual).
IMO the proposition is little fishy, but its not totally without merit and imo deserves investigation. If we are all worried about our jobs, even via building custom for sale software, there is likely something there that may obviate the need at least for end user applications. Again, im deeply skeptical, but it is interesting.
> Being able to coat efficiently run frontier models is i think, not a high priced endeavor for an org
Running proprietary model would make you subject to whatever ToS the LLM companies choose on a particular day, and what you can produce with them, which circles back to the raison d'etre for the GPL and GNU.
Until all software copyright is dead and buried, there is no need for copyleft to change tack. Otherwise there rising tide may rise high enough to drown GPL, but not proprietary software.
Open source is easier to counterfeit/license-launder/re-implement using LLMs because source code is much lower-hanging fruit, and is understood by more people than closed-source assembly.
How close are we to good enough and who's working on that? I would be interested in supporting that work; to my mind, many of the real objections to LLMs are diminished if we can make them small and cheap enough to run in the home (and, perhaps, trained with distributed shared resources, although the training problem is the harder one).
Good question. It seems like most of the tech world is perfectly happy to be sharecroppers on the Big AI farms. I guess that's not quite the right analogy, since they're doing their own things with it; just that at the end of the day, the tool they're building everything on is owned by someone else.
Copyleft is a mirror of copyright, not a way to fight copyright. It grants rights to the consumer where copyright grants rights to the creator. Importantly, it gives the end-user the right to modify the software running on their devices.
Unfortunately, there are cases where you simply can't just "re-implement" something. E.g., because doing so requires access to restricted tools, keys, or proprietary specifications.
"So, I looked for a way to stop that from happening. The method I came up with is called “copyleft.” It's called copyleft because it's sort of like taking copyright and flipping it over. [Laughter] Legally, copyleft works based on copyright. We use the existing copyright law, but we use it to achieve a very different goal."
"very different goal" isn't the same as "fundamentally destroying copyright"
the very different goal include to protect public code to stay public, be properly attributed, prevent companies from just "sizing" , motivate other to make their code public too etc.
and even if his goals where not like that, it wouldn't make a difference as this is what many people try to archive with using such licenses
this kind of AI usage is very much not in line with this goals,
and in general way cheaper to do software cloning isn't sufficient to fix many of the issues the FOSS movement tried to fix, especially not when looking at the current ecosystem most people are interacting with (i.e. Phones)
---
("sizing"): As in the typical MS embrace, extend and extinguish strategy of first embracing the code then giving it proprietary but available extensions/changes/bug fixes/security patches to then make them no longer available if you don't pay them/play by their rules.
---
Through in the end using AI as a "fancy complicated" photocopier for code is as much removing copyright as using a photocopier for code would. It doesn't matter if you use the photocopier blind folded and never looked at the thing you copied.
That’s not a rebuttal of the OP’s point. None of that says anything about fighting copyright. It literally says he flipped it which is wha the OP said when they said it’s a mirror.
> AI is eroding copyright, so there may no longer be a need for the GPL. GNU should stop and rethink its stance, chuck away the GPL as the main tool to fight evil software corporations and embrace LLM as the main weapon.
Is this LLM thing freely available or is it owned and controlled by these companies? Are we going to rent the tools to fight "evil software corporations"?
There already are LLMs with open weights that are better at code than state of the art closed source models from a year ago. For now, you most people may have to rent the hardware to run those models, since it's too expensive for most people to own something that can run inference on one trillion parameters, but I wouldn't consider LLMs to be controlled by "evil software corporations" at this point.
With the release of GLM-5, I would say that they are pretty much almost as good. Basically 90% as good as Opus 4.6 on most tasks for 20% of inference cost, and open weights.
Until there is a capable open source open weight AI that is easily hostable by an average person - no, we still have a long way to go. You aren't going to have software freedom when the tool that enables it is controlled by a handful of powerful tech companies.
> we'll see that it was an attempt to fight copyrights with copyrights
it's not that simple
yes, GPLs origins have the idea of "everyone should be able to use"
but it also is about attribution the original author
and making sure people can't just de-facto "size public goods"
the kind of AI usage is removing attribution and is often sizing public goods in a way far worse then most companies which just ignored the license did
so today there is more need then ever in the last few decades for GPL like licenses
Its purpose "if you run the software you should be able to inspect and modify that software, and to share those modifications with your peers" not explicitly resist copyright. Yes copyright is bad in that it often prevents one from doing that, but it is not the purpose of the GPL to dismantle copyright.
Reducing it to "well you can clone the proprietary software you're forced to use by LLM" is really missing the soul of the GPL.
Just because something is copyleft doesn't mean the person who gave you the binary you're using has to supply you with the code the used to build it. That's what the GPL does.
I agree with almost all of that, except the part about GNU changing their stance. I think GNU should stay true and consistent, if for no other reason than to not make many of their supporters who aren't on board with AI feel betrayed and have GNUs legacy soured. If the cause of LLMs conquering proprietary software needs an organization to champion it, let that be a new organization, not GNU.
That's naive. Copyright doesn't just apply to software. There already have been countless lawsuits about copying music long before the term "open source" was invented. No, changing the lyrics a bit doesn't circumvent copyright. Nor does translating a Stephen King novel to German and switching the names of the places and characters.
A court ordered the first Nosferatu movie to be destroyed because it had too many similarities to Dracula. Despite the fact that the movie makes rather large deviations from the original.
If Claude was indeed asked to reimplement the existing codebase, just in Rust and a bit optimized, that could well be a copyright violation. Just like rephrasing A Song ot Ice and Fire a bit, and switching to a different language, doesn't remove its copyright.
Claude was asked to implement a public API, not an entire codebase. The definition of a public API is largely functional; even in an unusually complex case like the Java standard facilities (which are unusually creative even in the structure and organization of the API itself) the reimplementation by Google was found to be fair use.
> Claude was asked to implement a public API, not an entire codebase.
Allegedly. There have been several people who doubted this story. So how to find out who is right? Well, just let Claude compare the sources. Coincidentally, Claude Opus 4.6 doesn't just score 75.6% on SWE-bench Verified but also 90.2% on BigLaw Bench.
It's like our copyright lawyer is conveniently also a developer. And possibly identical to the AI that carried out the rewrite/reimplemention in question in the first place.
> Just like rephrasing A Song ot Ice and Fire a bit, and switching to a different language, doesn't remove its copyright.
There is some precedent for this, e.g. Alchemised is a recent best seller that had just enough changed from its Harry Potter fan fiction source in order to avoid copyright infringement: https://en.wikipedia.org/wiki/Alchemised
(I avoided the term “remove copyright” here because the new work is still under copyright, just not Harry Potter - related copyright.)
So not only are we moving goalposts here, but we've decided the GNU team should join the other team? I don't understand how GNU would see mass model LLM training as anything but the most flagrant violations of their ethos. LLM labs, in their view, would be among the most evil software corporations to have ever existed.
While I personally agree with you, Richard Stallman (the creator of the GPL) does not. He has always advocated in favor of strong copyright protection, because the foundation of the GPL is the monopoly power granted by copyright. The problem that the GPL is intended to solve is proprietary software.
Generative models (AI) are not really eroding copyright. They are calling its bluff. The very notion of intellectual property depends on a property line: some arbitrary boundary where the property begins and ends. Generative models blur that line, making it impractical to distinguish which property belongs to whom.
Ironically, these models are made by giant monopolistic corporations whose wealth is quite literally a market valuation (stock price) of their copyrights! If generative models ever become good enough to reimplement CUDA, what value will NVIDIA have left?
The reality is that generative models are nowhere near good enough to actually call the bluff. Copyright is still the winning hand, and that is likely to continue, particularly while IP holders are the primary authors of law.
---
This whole situation is missing the forest for the trees. Intellectual Property is bullshit. A system predicated on monopoly power can only result in consolidated wealth driving the consolidation of power; which is precisely what has happened. The words "starving artist" ring every bit as familiar today as any time in history. Copyright has utterly failed the very goals it was explicitly written with.
It isn't the GPL that needs changing. So long as a system of copyright rules the land, copyleft is the best way to participate. What we really need is a cohesive political movement against monopoly power; one that isn't conveniently ignorant of copyright as its most significant source.
I think the missing thing here is that the license violation already happened. Most of the big models trained on data in a manner that violated terms of service. We'll need a court case but I think it's extremely reasonable to consider any model trained on GPL code to be infected with open licensing requirements.
You might wish that were true, but there are very strong arguments it's not. Training on copyleft licensed code is not a license violation. Any more than a person reading it is. In copyright terms, it's such an extreme transformative use that copyright no longer applies. It's fair use.
But agreed that we're waiting for a court case to confirm that. Although really, the main questions for any court cases are not going to be around the principle of fair use itself or whether training is transformative enough (it obviously is), but rather on the specifics:
1) Was any copyrighted material acquired legally (not applicable here), and
2) Is the LLM always providing a unique expression (e.g. not regurgitating books or libraries verbatim)
And in this particular case, they confirmed that the new implementation is 98.7% unique.
Transformative is not the only component of determining fair use, there’s also the economic displacement aspect. If you’re doing a book report and include portions of the original (or provide an interface for viewing portions à la Google Books) you aren’t a threat to the original authors ability to make a living.
If you’ve used copyrighted books and turned them into a free write-a-book machine, you are suddenly using the authors own works against them, in a way that a judge might rule is not very fair.
“ Effect of the use upon the potential market for or value of the copyrighted work: Here, courts review whether, and to what extent, the unlicensed use harms the existing or future market for the copyright owner’s original work. In assessing this factor, courts consider whether the use is hurting the current market for the original work (for example, by displacing sales of the original) and/or whether the use could cause substantial harm if it were to become widespread.”
A human reading a unit of work is not a “copy”. I’m pretty sure our legal systems agree that thought or sight is not copying something.
Training an LLM inherently requires making a copy of the work. Even the initial act of loading it from the internet and copying it into memory to then train the LLM is a copy that can be governed by its license and copyright law
I think you are confusing two different meanings of the word ‘copy’. The fact that a computer loads it into memory does not make it automatically a ‘copy’ in the copyright sense.
> Training an LLM inherently requires making a copy of the work.
But that's not relevant here. Because the copyleft license does not prohibit that (and it's not even clear that any license can prohibit it, as courts may confirm it's fair use, as most people are currently assuming). That's why I noted under (1) that it's not applicable here.
It's absolutely prohibited to copy and redistribute for commercial purposes materials that you're unlicensed to do so with. This isn't an issue when it comes to the copy-left scenario (though it may potentially enforce transitive licensing requirements on the copier that LLM runners don't want to follow) but it is a huge issue that has come up with LLM training.
LLM training involves ingesting works (in a potentially transformative process) and partially reproduce them - that's a generally restricted action when it comes to licensing.
> It's absolutely prohibited to copy and redistribute for commercial purposes materials that you're unlicensed to do so with.
Sure, but that's not what LLM's generally do, and it's certainly not what they're intended to do.
The LLM companies, and many other people, argue that training falls under fair use. One element of fair use is whether the purpose/character is sufficiently transformative, and transforming texts into weights without even a remote 1-1 correspondence is the transformation.
And this is why LLM companies ensure that partial reproduction doesn't happen during LLM usage, using a kind of copyrighted-text filter as a last check in case anything would unintentionally get through. (And it doesn't even tend to occur in the first place, except when the LLM is trained on a bunch of copies of the same text.)
Yea, at the end of the day a big part of this question comes down to whether that copying is fair use and that is an open question with the transformative nature being the primary point in favor of the LLM. But it is copying from some works to another - if it doesn't have some fair use exception it is absolutely violating the licensing of most of the training data. It's a bit different from previous settled case law because it's copying so little from so many billions of different things. I think blocking reproduction is wise by LLM companies for PR purposes but it doesn't guarantee that training is a license exempted activity.
Yup. Of course it's copying. But all expectations are that courts will rule that fair use allows such copying, because of the nature of the transformation.
> Training on copyleft licensed code is not a license violation. Any more than a person reading it is.
Some might hold that we've granted persons certain exemptions, on account of them being persons. We do not have to grant machines the same.
> In copyright terms, it's such an extreme transformative use that copyright no longer applies.
Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim? Sure, it can also produce extremely transformed versions, but is that really relevant if it holds within it enough information for a (near-)verbatim reproduction?
>Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim? Sure, it can also produce extremely transformed versions, but is that really relevant if it holds within it enough information for a (near-)verbatim reproduction?
I feel as though, from an information-theoretic standpoint, it can't be possible that an LLM (which is almost certainly <1 TB big) can contain any substantial verbatim portion of its training corpus, which includes audio, images, and videos.
No we don't have to, but so far we do, because that's the most legally consistent. If you want to change that, you're going to need to pass new laws that may wind up radically redefining intellectual property.
> Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim?
Of course it has, if the transformation is extreme, as it appears to be here. If I memorize the lyrics to a bunch of love songs, and then write my own love song where every line is new, nobody's going to successfully sue me just because I can sing a bunch of other songs from memory.
Also, it's not even remotely clear that the LLM can produce the training data near-verbatim. Generally it can't, unless it's something that it's been trained on with high levels of repetition.
> you're going to need to pass new laws that may wind up radically redefining intellectual property
You're correct that this is one route to resolving the situation, but I think it's reasonable to lean more strongly into the original intent of intellectual property laws to defend creative works as a manner to sustain yourself that would draw a pretty clear distinction between human creativity and reuse and LLMs.
> into the original intent of intellectual property laws to defend creative works as a manner to sustain yourself
But you're missing the other half of copyright law, which is the original intent to promote the public good.
That's why fair use exists, for the public good. And that's why the main legal argument behind LLM training is fair use -- that the resulting product doesn't compete directly with the originals, and is in the public good.
In other words, if you write an autobiography, you're not losing significant sales because people are asking an LLM about your life.
The big difference between people reading code and LLMs reading code is that people have legal liability and LLMs do not. You can't sue an LLM for copyright infringement, and it's almost impossible for users to tell when it happens.
BTW in 2023 I watched ChatGPT spit out hundreds of lines of F# verbatim from my own GitHub. A lot of people had this experience with GitHub Copilot. "98.7% unique" is still a lot of infringement.
If you comission art from an artist who paints a modified copy of Warhol's work, the artist is liable (even if you keep that work private, for personal use).
If you commission it from OpenAI (by sending a query to their ChatGPT API), by your argument, you are the person liable — and OpenAI is off the hook even if that work is distributed further.
I'm not going to argue about the merits of creativity here, or that someone putting a prompt into ChatGPT considers themselves an artist.
That's irrelevant. The work is created on OpenAI servers, by the LLMs hosted there, and is then distributed to whoever wrote the prompt.
Models run locally are distributed by whoever trained them.
If you train a model on whatever data you legally have access to, and produce something for yourself, it's one thing.
Distribution is where things start to get different.
I agree there has to be a court case about it. I think the current argument, however, is that it is transformative, and therefore falls under fair use.
You don’t get to simply claim fair use based on how transformative your derivative work is.
“””
Section 107 calls for consideration of the following four factors in evaluating a question of fair use:
Purpose and character of the use, including whether the use is of a commercial nature or is for nonprofit educational purposes: Courts look at how the party claiming fair use is using the copyrighted work, and are more likely to find that nonprofit educational and noncommercial uses are fair. This does not mean, however, that all nonprofit education and noncommercial uses are fair and all commercial uses are not fair; instead, courts will balance the purpose and character of the use against the other factors below. Additionally, “transformative” uses are more likely to be considered fair. Transformative uses are those that add something new, with a further purpose or different character, and do not substitute for the original use of the work.
Nature of the copyrighted work: This factor analyzes the degree to which the work that was used relates to copyright’s purpose of encouraging creative expression. Thus, using a more creative or imaginative work (such as a novel, movie, or song) is less likely to support a claim of a fair use than using a factual work (such as a technical article or news item). In addition, use of an unpublished work is less likely to be considered fair.
Amount and substantiality of the portion used in relation to the copyrighted work as a whole: Under this factor, courts look at both the quantity and quality of the copyrighted material that was used. If the use includes a large portion of the copyrighted work, fair use is less likely to be found; if the use employs only a small amount of copyrighted material, fair use is more likely. That said, some courts have found use of an entire work to be fair under certain circumstances. And in other contexts, using even a small amount of a copyrighted work was determined not to be fair because the selection was an important part—or the “heart”—of the work.
Effect of the use upon the potential market for or value of the copyrighted work: Here, courts review whether, and to what extent, the unlicensed use harms the existing or future market for the copyright owner’s original work. In assessing this factor, courts consider whether the use is hurting the current market for the original work (for example, by displacing sales of the original) and/or whether the use could cause substantial harm if it were to become widespread.
“””
Yea, a finding that training is transformative would be pretty significant and it's likely that the precedent of thumbnail creation being deemed transformative would likely steer us towards such a finding. Transformative is always a hard thing to bank on because it is such a nebulous and judgement based call. There are excellent examples of how precise and gritty this can get in audio sampling.
Didn't know about thumbnails being fair use. In that case, I just don't see an argument that genAI training on source code is less transformative than thumbnails.
> He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch.
From GPL2:
> The source code for a work means the preferred form of the work for making modifications to it. For an executable work, complete source code means all the source code for all modules it contains, plus any associated interface definition files, plus the scripts used to control compilation and installation of the executable.
Is a project's test suite not considered part of its source code? When I make modifications to a project, its test cases are very much a part of that process.
If the test suite is part of this library's source code, and Claude was fed the test suite or interface definition files, is the output not considered a work based on the library under the terms of LGPL 2.1?
Legally, using the tests to help create the reimplementation is fine.
However, it seems possible you can't redistribute the same tests under the MIT license. So the reimplementation MIT distribution could need to be source code only, not source code plus tests. Or, the tests can be distributed in parallel but still under LGPL, not MIT. It doesn't really matter since compiled software won't be including the tests anyways.
Sorry, I misspoke. Transformation is what makes the LLM itself legal -- its training data is sufficiently transformed into weights.
And so, a work being sufficiently transformative is one way in which copyright no longer applies, but that's not the case here specifically. The specific case here is essentially just a clean-room reimplementation (though technically less "clean", but still presumably the same legally). But the end result is still a completely different expression of underlying non-copyrightable ideas.
And in both cases, it doesn't matter what the original license was. If a resulting work is sufficiently transformative or a reimplementation, copyright no longer applies, so the license no longer applies.
The library's test suite and interfaces were apparently used directly, not transformed. If either of those are considered part of the library's source code, as the license's wording seems to suggest, then I think output from their use could be considered a work based on the library as defined in the license.
Google LLC v Oracle America assumed (though didn't establish) that API's are copyrightable... BUT that developing against them falls under fair use, as long as the function implementations are independent.
Test suites are again generally considered copyrightable... but the behavior being tested is not.
So no, it's not considered to be a work based on the library. This seems pretty clear-cut in US law by now.
Also, the LGPL text doesn't say "work based on the library". It says "If you modify a copy of the Library", and this is not a "combined work" either. And the whole point is that this is not a modified copy -- it's a reimplementation.
In theory, a license could be written to prevent running its tests from being run against software not derived from the original, i.e. clean-room reimplementations. In practice, it remains dubious whether any court would uphold that. And it would also be trivial to then get around it, by taking advantage of fair use to re-implement the tests in e.g. plain English (or any specification language), and then re-implementing those back into new test code. Because again, test behaviors are not copyrightable.
Google vs Oracle ruled that APIs fall under copyright (the contrary was thought before). However, it was ruled that, in that specific case, fair use applied, because of interoperability concerns. That's the important part of this case: fair use is never automatic, it is assessed case by case.
Regarding chardet, I'm not sure "I wanted to circumvent the license" is a good way to argue fair use.
> Blanchard's account is that he never looked at the existing source code directly. He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch
This feels sort of like saying "I just blindly threw paint at that canvas on the wall and it came out in the shape of Mickey Mouse, and so it can't be copyright infringement because it was created without the use of my knowledge of Micky Mouse"
Blanchard is, of course, familiar with the source code, he's been its maintainer for years. The premise is that he prompted Claude to reimplement it, without using his own knowledge of it to direct or steer.
> Blanchard is, of course, familiar with the source code, he's been its maintainer for years.
I would argue it's irrelevant if they looked or didn't look at the code. As well as weather he was or wasn't familiar with it.
What matters is, that they feed to original code into a tool which they setup to make a copy of it. How that tool works doesn't really matter. Neither does it make a difference if you obfuscate that it's an copy.
If I blindfold myself when making copies of books with a book scanner + printer I'm still engaging in copyright infringement.
If AI is a tool, that should hold.
If it isn't "just" a tool, then it did engage in copyright infringement (as it created the new output side by side with the original) in the same way an employee might do so on command of their boss. Which still makes the boss/company liable for copyright infringement and in general just because you weren't the one who created an infringing product doesn't mean you aren't more or less as liable of distributing it, as if you had done so.
>that they feed to original code into a tool which they setup to make a copy of it
Well, no. They fed the spec (test cases, etc) into a tool which made a new program matching the spec. This is not a copy of the original code.
But also this feels like arguing over the color of the iceberg while the titanic sinks. If you have a tool that can make code to spec, what is the value in source code anymore? Even if your app is closed-source, you can just tell claude to write new code that does the same thing.
Blanchard fed the spec to the tool, and Anthropic fed the code to the tool, so Blanchard didn't do anything wrong, and Anthropic didn't do anything wrong. Nothing to see here.
Everyone writes as if he just fed the spec and tests to Claude Code. Ignoring for now that the tests are under LGPL as well, the commit history shows that this has been done with two weeks of steering Claude Code towards the desired output. At every one of these interactions, the maintainer used his deep knowledge of the chardet codebase to steer Claude.
What does derivative mean here? Because IMO it means that the existing work was used as input. So if you used a LLM and it was trained on the existing work, that's a derivative work. If you rot13 encode something as input, so you can't personally read it, and then a device decides to rot13 on it again and output it, that's a derivative work.
In order for it to be creatively derivative you would need to copy the structure, logic, organization, and sequence of operations not just reimplement the functionality. It is pretty clear in this case that wasn't done.
As a cynical person I assume all the frontier LLMs were trained on datasets that include every open source project, but as a thought experiment, if an LLM was trained on a dataset that included every open source project _execept_ chardet, do you think said LLM would still be able to easily implement something very similar?
Of course, the problem with this interpretation is that all modern LLMs are derivatives from huge amounts of text under completely different licenses, including "All rights reserved", and therefore can not be used for any purpose.
I'm not sure how you square the circle of "it's alright to use the LLM to write code, unless the code is a rewrite of an open source project to change its license".
> Of course, the problem with this interpretation is that all modern LLMs are derivatives from huge amounts of text under completely different licenses, including "All rights reserved", and therefore can not be used for any purpose.
> I'm not sure how you square the circle of "it's alright to use the LLM to write code
You seem like you're on the cusp of stating the obvious correct conclusion: it isn't.
LLMs do not encode nor encrypt their training data. The fact they can recite training data is a defect not a default. You can understand this more simply by calculating the model size as an inverse of a fantasy compression algorithm that is 50% better than SOTA. You'll find you'd still be missing 80-90% of the training data even if it were as much of a stochastic parrot as you may be implying. The outputs of AI are not derivative just because they saw training data including the original library.
Then onto prompting: 'He fed only the API and (his) test suite to Claude'
This is Google v Oracle all over again - are APIs copyrightable?
> LLMs do not encode nor encrypt their training data. The fact they can recite training data is a defect not a default.
About this specific point, it is unclear how much of a defect memorization actually is - there are also reasons to see it as necessary for effective learning. This link explains it well:
> This is Google v Oracle all over again - are APIs copyrightable?
Yes this is the best way to ask the question. If I take a public facing API and reimplement everything, whether it's by human or machine, it should be sufficient. After all, that's what Google did, and it's not like their engineers never read a single line of the Java source code. Even in "clean room" implementations, a human might still have remembered or recalled a previous implementation of some function they had encountered before.
I find the "compression" argument not very strong, both because copyright still applies to (very) lossy codecs (e.g. your 16kbps Opus file of Thriller infringes, even if the original 192khz/32bit wav file was 12,000kbps), and because copyright still applies to transformed derivative works (a tiny midi file of Thriller might still be enough for the Jackson's label to get you)
Copyright protects even very abstract aspects of human creative expression, not just the specific form in which it is originally expressed. If you translate a book into another language, or turn it into a silent movie, none of the actual text may survive, but the story itself remains covered by the original copyright.
So when you clone the behavior of a program like chardet without referencing the original source code except by executing it to make sure your clone produces exactly the same output, you may still be infringing its copyright if that output reflects creative choices made in the design of chardet that aren't fully determined by the functional purpose of the program.
If you pirate a movie and reencode it, does that apply as well? You can still watch the movie and it is “obviously” the same movie, even though the bytes are completely different. Here you can use the program and it is, to the user, also the same.
So, let's say that rather than actually touching any copyrighted material, a human merely tells an AI about how to go onto the internet and find copyrighted material, download it, and ingest it for training. The AI, fully autonomously, does so, and after training itself on the material deletes it so no human ever downloads, consumes, or shares it.
If we are saying AI is "more than a tool", which seems to be the case courts are leaning since they've ruled AI output without direct human involvement is not copyrightable[0], then the above seems like it would be entirely legal.
Someone would likely get prosecuted if they instructed AI agent to run say a pump and dump scheme...
Even if the final output doesn't have copyright protection it might still be copyright violation. I think it could be reasonable to have work that itself violates copyright when distributed even if it does not have copy right itself.
I just don't see how it's relevant whether he did look or didn't. In my opinion, it's not just legally valid to make a re-implementation of something if you've seen the code as long as it doesn't copy expressive elements. I think it's also ethically fine as well to use source code as a reference for re-implementing something as long as it doesn't turn into an exact translation.
It's actually not legally fine, or at least it's extremely dangerous. Projects that re-implement APIs presented by extremely litigious companies specifically do not allow people who, for instance, have seen the proprietary source code to then work on the project.
I don't think fear or legal action makes it illegal.
If I know it is legal to make a turn at a red light. And I know a court will uphold that I was in the right but a police officer will fine me regardless and I would need to go to actually pursue some legal remedy I'm unlikely to do it regardless of whether it is legal because it is expensive, if not in money but time.
In the case of copyright lawsuits they are notoriously expensive and long so even if a court would eventually deem it fine, why take the chance.
That's my point. It's dangerous and there are sharks in the water. That sounds like you're not going to have a good time if you do the described approach to someone who might assert you're infringing.
Right. The alternative is that we reward Dan for his 14 years of volunteer maintenance of a project... by banning him from working on anything similar under a different license for the rest of his life.
Ignoring the legal or ethical concerns. Let’s say we live in a world where the cost of copying code is so close to zero that it’s indistinguishable from a world without copyright.
Anything you put out can and will be used by whatever giant company wants to use it with no attribution whatsoever.
Doesn’t that massively reduce the incentive to release the source of anything ever?
No, because (most) people don't work on OSS for vanity, they do it to help other people, whether it's individuals or groups of individuals, ie corporations.
It's the same question as, if an AI can generate "art", or photographers can capture a scene better than any (realistic) painter, then will people still create art? Obviously yes, and we see it of course after Stable Diffusion was released three years ago, people are still creating.
I don’t know what a world without copyright does to corporate sponsored open source. It certainly reduces it because there are many corporate sponsored projects that monetize through dual licensing. My guess is in a world where you can’t even guarantee attribution, it’s much harder to convince your boss to let you open source a project in the first place.
So ignoring people who are being paid by corporations directly to work on open source, in my experience the vast majority of contributors expect to be able to monetize their work eventually in a way that requires attribution. And out of the small number who don’t expect a monetary return of any kind, a still smaller number don’t expect recognition.
If this weren’t the case you’d see a much larger amount of anonymous contributions. There are people who anonymously donate to charity. The vast majority want some kind of recognition.
Obviously we still see art, if you greatly reduce the monetary benefit to producing art, you’ll see a lot less of it. This is especially true of non trivial open source software that unlike static artwork requires continual maintenance.
Yes, and it reduces the incentives to release binaries too. Such a world will be populated by almost entirely SaaS, which can still compete on freedom.
Oracle had it's day in court with Google over the Java APIs. Reimplementing APIs can be done without copyright infringement, but Oracle must have tried to find real infringement during discovery.
In this case, we could theoretically prove that the new chardet is a clean reimplementation. Blanchard can provide all of the prompts necessary to re-implement again, and for the cost of the tokens anyone can reproduce the results.
>This feels sort of like saying "I just blindly threw paint at that canvas on the wall and it came out in the shape of Mickey Mouse, and so it can't be copyright infringement because it was created without the use of my knowledge of Micky Mouse"
IANAL, but that analogy wouldn't work because Mickey Mouse is a trademark, so it doesn't matter how it is created.
If you only stick to the API and ignore the implementation, it is not Mickey Mouse any more but a rodent. If it was just a clone it wouldn't be 50x as fast. Nevertheless, APIs apparently can be copyrightable. I generally disagree with this; it's how PC compatibles took off, giving consumers better options.
The article is poorly written. Blanchard was a chardet maintainer for years. Of course he had looked at it's code!
What he claimed, and what was interesting, was that Claude didn't look at the code, only the API and the test suite. The new implementation is all Claude. And the implementation is different enough to be considered original, completely different structure, design, and hey, a 48x improvement in performance! It's just API-compatible with the original. Which as per the Google Vs oracle 2021 decision is to be considered fair use.
What if we said that generative AI output is simply not copyrightable. Anything an AI spits out would automatically be public domain, except in cases where the output directly infringes the rights of an existing work.
This would make it so relicensing with AI rewrites is essentially impossible unless your goal is to transition the work to be truly public domain.
I think this also helps somewhat with the ethical quandary of these models being trained on public data while contributing nothing of value back to the public, and disincentivize the production of slop for profit.
> No Copyright Protection for AI-Assisted Creations: Thaler v. Perlmutter
> A recent key judicial development on this topic occurred when the U.S. Supreme Court declined to review the case of Thaler v. Perlmutter on March 2, 2026, effectively upholding lower court rulings that AI-generated works lacking human authorship are not eligible for copyright protection under U.S. law
> > A recent key judicial development on this topic occurred when the U.S. Supreme Court declined to review the case of Thaler v. Perlmutter on March 2, 2026, effectively upholding lower court rulings that AI-generated works lacking human authorship are not eligible for copyright protection under U.S. law
This was AI summary? Those words were not in the article.
The courts said Thaler could not have copyright because he refused to list himself as an author.
In the corporate world, we've started using reimplementation as a way to access tooling that security won't authorize.
Sec has a deny by default policy. Eng has a use-more-AI policy. Any code written in-house is accepted by default. You can see where this is going.
We've been using AI to reimplement tooling that security won't approve. The incentives conspired in the worst outcome, yet here we are. If you want a different outcome, you need to create different incentives.
Not Invented Here's long, slow mutagenic march toward full antibiotic resistance continues apace.
There is a fundamental corpo-cognitive dissonance, to boot. If "AI" is cheap enough and good enough to implement security-relevant software from `git init` repeatedly, why isn't it also cheap enough and good enough to assess and approve the security of third-party software at pace with internal adoption? Is there some basis to believe LLMs' leverage on production differs from its leverage on analysis of existing code?
I largely agree with the author that AI can't just magically remove license agreements by rewriting code.
However, I take issue with his version of history:
>The history of the GPL is the history of licensing tools evolving in response to new forms of exploitation: GPLv2 to GPLv3, then AGPL.
GPLv3 set open source backwards: it wasn't an evolution to protect anything, it was a an overly paranoid failure. Don't believe me? Just count how many GPL3 vs. how many GPL2 projects have been started since GPL3 dropped.
Again, I'm very pro-OSS, but let's not pretend the community has always had a straight line of progress forward; some stuff is crazy Stallman stuff that set us back.
If Blanchard is claiming not to have been substantively involved in the creation of the new implementation of chardet (i.e. "Claude did it"), then the new implementation is machine generated, and in the USA cannot be copyright and thus cannot be licensed.
If he is claiming to have been somehow substantively "enough" involved to make the code copyrightable, then his own familiarity with the previous LGPL implementation makes the new one almost certainly a derivative of the original.
>then his own familiarity with the previous LGPL implementation makes the new one almost certainly a derivative of the original.
The "clean room rewrite" is just an extreme way to have a bulletproof shield against litigation. Not doing it that way doesn't automatically make all new code he writes derivative solely because he saw how the code worked previously.
If the clean room re-write was done entirely by Claude, then the result cannot be copyright in the USA, and thus there is no license at all.
And if he was in fact more involved (which he appears to deny) that it's a bit weak to say that someone with huge familiarity with chardet could choose to reimplement chardet without the result being derivative.
There's a difference between "I've read a LGPL code once, maybe I could do something similar" and "I've been reading this LGPL code for 12 years and now I'm going to do exactly the same thing".
I agree with the thrust of this article, that norms and what we perceive as good or desirable extend considerably beyond the minimum established by law.
But a point that was not made strongly, which highlights this even more, is that this goes in every direction.
If this kind of reimplementation is legal, then I can take any permissive OSS and rebuild it as proprietary. I can take any proprietary software and rebuild it as permissive. I can take any proprietary software and rebuild it as my own proprietary software.
Either the law needs to catch up and prevent this kind of behavior, or we're going to enter an effectively post-copyright world with respect to software. Which ISN'T GOOD, because that will disincentivize any sort of open license at all, and companies will start protecting/obfuscating their APIs like trade secrets.
Companies can take open-source software and make a proprietary reimplementation. You can't take a proprietary software and make an open source GPL version.
I am absolutely certain that if you tried you would be sued to oblivion. But big company screwing up open source is not even news anymore. In fact I (still) believe that the fact that even though LLMs were trained on tons of GPL and AGPL or even unlicensed software it's considered ok to use LLM code in proprietary projects is example of just that.
From a strictly legal perspective the two are equivalent. The fact that there are structural injustices in the system is true, but that's not a question that any answer to "what should be legal" can fix.
I've been thinking this for over two years, that's why I stopped contributing to open source at that time - my work was only gonna be exploited to make rich people richer regardless of the license.
Crazy that only now we're seeing a bunch of articles coming to the same conclusion now.
I think copyright should still apply, but if it doesn't, we need new laws - ones which protect all human work, creative or not. Laws should serve and protect people, not algorithms and not corporations "owning" those algorithms.
I put owning in quotes because ownership should go to the people who did the work.
Buying/selling ownership of both companies and people's work should be illegal just like buying/selling whole humans is. Even if it took thousands of years to get here.
Money should not buy certain things because this is the root cause of inequality. Rich people are not getting richer at a faster rate by being more productive than everyone else but by "owning" other people's work and using it as leverage to extract even more from others.
Maybe LLM and mass unemployment of white collar workers will be the wakeup call needed for a reform. Or revolution.
Last time this happened was during the second industrial revolution and that's how communism got popular. We should do better this time because this is the last revolution which might be possible.
Wow, it feels like this argument rewired my brain.
When I first read about the chardet situation, I was conflicted but largely sided on the legal permissibility side of things. Uncomfortably I couldn't really fault the vibers; I guess I'm just liberal at heart.
The argument from the commons has really invoked my belief in the inherent morality of a public good. Something being "impermissible" sounds bad until you realize that otherwise the arrow of public knowledge suddenly points backwards.
Seeing this example play out in real life has had retroactive effects on my previously BSD-aligned brain. Even though the argument itself may have been presented before, I now understand the morals that a GPL license text underpins better.
FWIW I like to explain it to folks like this: ignore all of your moral baggage around licensing and just focus on the fact that licensing is a legal tool of art that pretty much only becomes relevant in the context of threatening lawsuits.
BSD-type stuff is very simple because it says "here is this stuff. you can use it as long as you promise not to sue me. I promise not to sue you too."
Very simple.
GPL-type stuff is intrinsically more complex because it's trying to use the threatening power of lawsuits, to reduce overall IP lawsuits. So it has to say "Here is this stuff. You can use it as long as you promise not to sue me. I am only going to sue you, if you start pretending like you have the right to sue other folks over this stuff or anything you derive from it. You don't have the right to sue others for it, I made it, so please stop pretending and let's stop suing each other over this sort of thing."
Getting the entire legal nuance around that sort of counterfactual "I will only sue you if you try to pretend that you can sue others" is why they're more complex. And the simplest copyleft licenses like the Mozilla Public License have a very rigid notion of what "the software" is, so like for MPL it's "this file is gonna never be used in a lawsuit, you can edit it ONLY as long as you agree that this file must never be used by you to sue someone else, if you try to mutate it in a way that lets you sue someone else then that's against our agreement and we reserve the right to sue you."
Whereas for GPL it's actually kind of nebulous what "the software" is -- everything that feeds into the eventual compiled binary, basically -- and so the license itself needs to be a little bit airy-fairy, "let's first talk about what conveying the software means...", in various ways.
The interesting thing here is that as far as the courts are initially ruling, these from-scratch reimplementations are not human works and therefore are not copyrightable, which makes them all kind of public domain. Slapping the MIT license on it was an overstep. If that's how things go then Free Software has actually won its greatest sweep with LLM ubiquity.
This is only worth arguing about because software has value. Putting this in context of a world where the cost of writing code is trending to 0, there are two obvious futures:
1. The cost continues to trend to 0, and _all_ software loses value and becomes immediately replaceable. In this world, proprietary, copyleft and permissive licenses do not matter, as I can simply have my AI reimplement whatever I want and not distribute it at all.
2. The coding cost reduction is all some temporary mirage, to be ended soon by drying VC money/rising inference costs, regulatory barriers, etc. In that world we should be reimplementing everything we can as copyleft while the inferencing is good.
There’s an other option. The cost of copying existing software trends to 0, but the cost of writing new software stays far enough above 0 that it is still relatively expensive.
There was a recent ruling that LLM output is inherently public domain (presumably unless it infringes some existing copyright). In which case it's not possible to use them to "reimplement everything we can as copyleft".
it's more complicated, the ruling was that AI can't be an author and the thing in question is (de-facto) public domain because it has no author in context of the "dev" claim it was fully build by AI
but AI assisted code has an author and claiming it's AI assisted even if it is fully AI build is trivial (if you don't make it public that you didn't do anything)
also some countries have laws which treat it like a tool in the sense that the one who used it is the author by default AFIK
The article is proceeding from the premise that a reimplementation is legal (but evil). To help my understanding of your comment, do you mean:
1. An LLM recreating a piece of software violates its copyright and is illegal, in which case LLM output can never be legally used because someone somewhere probably has a copyright on some portion of any software that an LLM could write.
2. You read my example as "copying a project without distributing it", vs. "having an LLM write the same functionality just for me"
There will always be cost though. Even if perfect code is getting one-shotted out, that is constantly maintained and adapted to changing conditions and technology, it simply can't stay at 0 forever because one day the power is surely going to go out!
More and more I am drawn to these kinds of ideas lately, perhaps as a kind of ethical sidestep, but still:
It's not going to solve any general issue here, but the one thing these freaks need that can't be generated by their models is energy, tons of it. So, the one thing I can do as an individual and in my (digital) community is work to be, in a word, self-sustainable. And depending on my company I guess, if I was a CEO I would hope I was wise enough to be thinking on the same lines.
Everyone is making beautiful mountains from paper and wire. I will just be happy to make a small dollhouse of stone, I think it will be worth it. How can we see not just at least some small-level of hubris otherwise?
I'm probably spitting in the wind, but stuff like this is why I removed all my hosted open source projects. I manage several niche projects that I have now converted to binary only releases (to almost no push back). It's niche enough that it's not very hard to get LLMs to output chunks of code that it managed to scrape before I took it offline. I don't see many people talking about this angle, but LLMs ripping off my work killed my open source efforts.
This take, which I've seen in a few different places now, seems 100% bonkers. A world where anyone can cheaply reimplement anyone else's software and use it on hardware of their own choosing in their own designs and for their own purposes is a free software utopia.
This isn't a problem, this is the goal. GNU was born when RMS couldn't use a printer the way he wanted because of an unmodifiable proprietary driver. That kind of thing just won't happen in the vibe coded future.
Surprised they don't mention Google LLC v. Oracle America, Inc. Seems a bit myopic to condone the general legality while arguing "you can only use it how I like it".
It also doesn't talk about the far more interesting philosophical queston. Does what Blanchard did cover ALL implementations from Claude? What if anyone did exactly what he did, feed it the test cases and say "re-implement from scratch", ostensibly one would expect the results to be largely similar (technically under the right conditions deterministically similar)
could you then fork the project under your own name and a commercial license? when you use an LLM like this, to basically do what anyone else could ask it to do how do you attach any license to it? Is it first come first serve?
If an agent is acting mostly on its own it feels like if you found a copy of Harry Potter in the fictional library of Babel, you didn't write it, just found it amongst the infinite library, but if you found it first could you block everyone else that stumbles on a near-identical copy elsewhere in the library? or does each found copy represent a "Re-implementation" that could be individually copyrighted?
It should be noted that the Rust community is also guilty of something similar. That is, porting old GPL programs, typically written in C, to Rust and relicensing them as MIT.
If an AI can license-wash open source software like this then the licenses become meaningless. Which is fascinating. Commercial software cloning that is simple enough for an average person to drive is next and the ultimate form of piracy, see an app for $10? Don’t fancy paying? Just ask ChatGPT for a clone. Future is going to be wild.
I take your point, but if the re-implementation looks the same, I would say it’s a form of copying. (Which I don’t think is a problem, I don’t think you should be able to own sequences of numbers.)
I think the direction we are going, the GPL is going to fade away. I think people will look at this like writing a book and claiming the ideas in the book cannot be copied. This debate is not that different from the ones going on in the music industry. I open sourced my latest software as Apache 2.0 after debating a lot about this. Unless the FSF wins in court in the next <=2-3 years, there is no coming back from this.
There's a Japanese version of that page, written in classical text writing direction, in columns. Which is cool. Makes me wonder, though - how readable is it with so many English loanwords which should be rotated sideways to fit into columns?
Total digression but yeah, that layout is stupid and the way those words are dropped in using Romaji makes no sense. That's not how Japanese people lay out pages on the web. In fact I don't think I've ever seen a Japanese web page laid out like a book like this, and in general I'd expect the English proper nouns and words that don't have obvious translations to get transliterated into Katakana. Smells like automatic conversion added by someone not really familiar with common practices for presenting Japanese on the web.
He also has a Korean vertical layout that lays out Latin-character words the same way. Is this common in Korea when vertical layout is used? The author seems to be Korean.
It seems that this chap didn't go and implement a new library, he reimplemented an existing one and became sole-controller of it. i.e. he seems to have taken its reputation, brand whatever you call it away from the contributors and entirely to himself. Their work of establishing it as a well known solution is no longer recognised.
So of course we feel that something wrong has happened even if it's not easy to put one's finger on it.
"Antirez closes his careful legal analysis as though it settles the matter. Ronacher acknowledges that “there is an obvious moral question here, but that isn't necessarily what I'm interested in.” Both pieces treat legal permissibility as a proxy for social legitimacy. "
This whole article is just complaining that other people didn't have the discussion he wanted.
Ronacher even acknowledged that it's a different discussion, and not one they were trying to have at the moment.
If you want to have it, have it. Don't blast others for not having it for you.
Having this discussion involves blasting others for not considering it. Consider the rest of the paragraph you quoted:
> But law only says what conduct it will not prevent—it does not certify that conduct as right. Aggressive tax minimization that never crosses into illegality may still be widely regarded as antisocial. A pharmaceutical company that legally acquires a patent on a long-generic drug and raises the price a hundredfold has not done something legal and therefore fine. Legality is a necessary condition; it is not a sufficient one.
Broadly speaking, the “freedom of users” is often protected by competition from competing alternatives. The GNU command line tools were replacements for system utilities. Linux was was a replacement for other Unix kernels. People chose to install them instead of proprietary alternatives. Was it due to ideology or lower cost or more features? All of the above. Different users have different motivations.
Copyleft could be seen as an attempt to give Free Software an edge in this competition for users, to counter the increased resources that proprietary systems can often draw on. I think success has been mixed. Sure, Linux won on the server. Open source won for libraries downloaded by language-specific package managers. But there’s a long tail of GPL apps that are not really all that appealing, compared to all the proprietary apps available from app stores.
But if reimplementing software is easy, there’s just going to be a lot more competition from both proprietary and open source software. Software that you can download for free that has better features and is more user-friendly is going to have an advantage.
With coding agents, it’s likely that you’ll be able to modify apps to your own needs more easily, too. Perhaps plugin systems and an AI that can write plugins for you will become the norm?
You can't put a copyright and MIT license on something you generated with AI. It is derived from the work of many unknown, uncredited authors.
Think about it; the license says that copies of the work must be reproduced with the copyright notice and licensing clauses intact. Why would anyone obey that, knowing it came from AI?
Countless instances of such licenses were ignored in the training data.
When learning is sufficiently atomized and recombined, creations cease to be "derived from" in a legal sense.
A lego sculpture is copyrighted. Lego blocks are not. The threshold between blocks and sculpture is not well-defined, but if an AI isn't prompted specifically to attempt to mimic an existing work, its output will be safely on the non-copyrighted side of things.
A derivative work is separately copyrightable, but redistribution needs permission from the original author too. Since that usually won't be granted or would be uneconomical, the derivative work can't usually be redistributed.
AI-produced material is inherently not copyrightable, but not because it's a derivative work.
Token prediction is a form of "learning" that is reinforced by the goal of reproducing the correct next token of the work, rather that acquiring ideas and concepts. For instance, given the prefix "Four score and seven years", the weights are adjusted until "ago" is correctly predicted, which is a fancy way of saying that it was stored in the model in a lossy way. The model "learned" that "ago" follows "four score and seven years" exactly the way your hard drive "learns" the audio and video frames of a movie when you download a .mp4 file.
I dispute the idea that token sequences reproduced from the model are not derived works.
I predict, no pun intended, that a time is coming when the idea that it's not a derived work will be challenged in mainstream law.
The slop merchants are getting a free ride for the time being.
As you said, it's lossy. Try it with any other distinctive but non-famous passage, and you won't get a correct prediction for the immediately following clause, much less for multiple sentences or paragraphs.
That's the case even when an LLM correctly identifies which book the prompted text is from. It still won't accurately continue on from some arbitrary passage. By the time you ask it to reproduce hundreds of words, you're into brand new book territory. Even when it's slop content, it's distinct slop.
The exceptions are cases where a significant number of humans would also know a particular quote from memory. Then, chances are, a frontier LLM will too.
You know how else you can reproduce a quote? Search for it on google, and search the resulting top hits; if it's a significant quote, multiple people have probably quoted it -- legally. You can also search a pirate library for the actual book, and search the book for the quote; while illegal, it's very simple to do, so unless you propose to make the free and open internet illegal, I'd suggest that banning LLMs for being "derivative work" creation engines is not so different from destroying the internet.
> I predict, no pun intended, that a time is coming when the idea that it's not a derived work will be challenged in mainstream law.
If judges have any sense whatsoever, LLM generations (without specific prompt crafting to mimic existing works) will be judged to not be derived works and therefore not be violating copyright, in the same sense that you can live and breathe Taylor Swift's music, create new music in the same style, and still not be violating copyright.
The Stability AI case, and how Judge Orrick deals with it, will be interesting and uninteresting at the same time. It deals primarily with the fact that after specific prompting, an image-generation AI can generate something fairly close to existing copyrighted images. That doesn't say anything more about whether LLMs are inherently producers of [only or primarily] derivative works, just as the fact that a human can violate copyright doesn't say anything about whether humans primarily or exclusively output derivative works.
More likely, perhaps, is that everything will be so infused with LLM output that copyright ceases to be relevant, or forces copyright law to be rewritten from the ground up.
Copyright requirements, even prior to LLMs, weren't well-specified. There's no objective threshold for how close something has to be to a previous work before the new one violates copyright. It's whatever a judge thinks, refering to the 4-factor test but ultimately making subjective judgements about each of those prongs. It's all a house of cards, and LLMs may just be what topples it.
This article is setting up a bit of a moving target. Legal vs legitimate is at least only a single vague question to be defined but then the target changes to “socially legitimate” defined only indirectly by way of example, like aggressive tax avoidance as “antisocial”— and while I tend to agree with that characterization my agreement is predicated on a layering of other principals.
The fundamental problem is that once you take something outside the realm of law and rule of law in its many facets as the legitimizing principal, you have to go a whole lot further to be coherent and consistent.
You can’t just leave things floating in a few ambiguous things you don’t like and feel “off” to you in some way- not if you’re trying to bring some clarity to your own thoughts, much less others. You don’t have to land on a conclusion either. By all means chew over things, but once you try to settle, things fall apart if you haven’t done the harder work of replacing the framework of law with that of another conceptual structure.
You need to at least be asking “to what ends? What purpose is served by the rule?” Otherwise you’re stuck in things where half the time you end up arguing backwards in ways that put purpose serving rules, the maintenance of the rule with justifications ever further afield pulled in when the rule is questioned and edge cases reached. If you’re asking, essentially, “is the spirit of the rule still there?” You’ve got to stop and fill in what that spirit is or you or people that want to control you or have an agenda will sweep in with their own language and fill the void to their own ends.
> If source code can now be generated from a specification, the specification is where the essential intellectual content of a GPL project resides. Blanchard's own claim—that he worked only from the test suite and API without reading the source—is, paradoxically, an argument for protecting that test suite and API specification under copyleft terms.
This is an interesting reversal in itself. If you make the specification protected under copyright, then the whole practice of clean room implementations is invalid.
Well, the license change sounds pretty strange, but to be honest if I were to use this software I would use it without adhering to the MIT. It's machine-created content which is not, in general, copyrightable. You can assert whatever license you want on such content, but I am not going to adhere to it. For example, I declare you may use the following under the Elastic License
> When GNU reimplemented the UNIX userspace, the vector ran from proprietary to free. Stallman was using the limits of copyright law to turn proprietary software into free software. […] The vector in the chardet case runs the other way.
That’s just your subjective opinion which many other people would disagree. I bet Armin Ronacher would agree that an MIT licensed library is even freer than an LGPL licensed library. To them, the vector is running from free to freer.
It also erodes copyright. A decent amount of commercial software can be AI cloned with no copyright violation.
A lot of SaaS too, especially if AI can run a simple deploy.
We might be approaching a huge deflationary catastrophe in the cost of a lot of software. It’s not a catastrophe for the consumer but it is for the industry.
It's clear that we're entering a new era of copyright _expectations_ (whether we get new _legislation_ is different), but for now realise this: the people like me who like copyleft can do this too. We can take software we like, point an agent at it, and tell it to make a new version with the AGPL3.0-or-later badge on the front.
no, it isn't. The point of the GPL is to grant users of the software four basic freedoms (run, study, modify and redistribute). There's no restriction to distribution per se, other than disallowing the removal of these freedoms to other users.
IMHO, the API and Test Suite, particularly the latter, define the contract of the functional definition of the software. It almost doesn't matter what that definition looks like so long as it conforms to the contract.
There was an issue where Google did something similar with the JVM, and ultimately it came down to whether or not Oracle owned the copyright to the header files containing the API. It went all the way to the US supreme court, and they ruled in Google's favour; finding that the API wasn't the implementation, and that the amount of shared code was so minimal as to be irrelevant.
They didn't anticipate that in less than half a decade we'd have technology that could _rapidly_ reimplement software given a strong functional definition and contract enforcing test suite.
without discussing copyright, I don't believe any of this is copied. Which I think should be the argument that actually matters.
I downloaded both 6.0 and 7.0 and based on only a light comparison of a few key files, nothing would suggest to me that 7.0 was copied from 6.0, especially for a 41x faster implementation.
It is a lot more organized and readable in my armature opinion, and the code is about 1/10th the size.
Not a lawyer, but my understanding is: In theory, copyright only protects the creative expression of source code; this is the point of the "clean room" dance, that you're keeping only the functional behavior (not protected by copyright). Patents are, of course, an entirely different can of worms. So using an LLM to strip all of the "creative expression" out of source code but create the same functionality feels like it could be equivalent enough.
I like the article's point of legal vs. legitimate here, though; copyright is actually something of a strange animal to use to protect source code, it was just the most convenient pre-existing framework to shove it in.
which is the actual relevant part: they didn't do that dance AFIK
AI is a tool, they set it up to make a non-verbatim copy of a program.
Then they feed it the original software (AFIK).
Which makes it a side by side copy, as in the original source was used as reference to create the new program. Which tend to be seen as derived work even if very different.
IMHO They would have to:
1. create a specification of the software _without looking at the source code_, i.e. by behavior observation (and an interface description). I.e. you give the AI access to running the program, but not to looking into the insides of it. I really don't think they did it as even with AI it's a huge pain as you normally can't just brute force all combinations of inputs and instead need to have a scientific model=>test=>refine loop (which AI can do, but can take long and get stuck, so you want it human assisted, and the human can't have inside knowledge about the program).
2. then generate a new program from specification, And only from it. No git history, no original source code access, no program access, no shared AI state or anything like that.
Also for the extra mile of legal risk avoidance do both human assisted and use unrelated 3rd parties without inside knowledge for both steps.
While this does majorly cut cost of a clean room approach, it still isn't cost free. And still is a legal mine field if done by a single person, especially if they have enough familiarity to potentially remember specific peaces of code verbatim.
Is your (1) description of clean room implementation, and (2) description of what was done, actually correct?
(1): my understanding was that a party _with access to copyrighted material_ made the functional spec, which was communicated to a party without access [1]. Under my understanding, theres no requirement for the authors of the functional spec to be 'clean'.
(2) Afaict, they limited the AI to access of just the functional spec and audited that it did not see the original source.
Edit: Not sure if sharing the 'test suite' matters, probably something for the courts in the unlikely event this ever gets there.
[1] Following the definition of clean room re implementation as it relates to US precedent, ie that described in the wikipedia page.
Well sure they didn't do the dance, but you don't have to do the dance. The reason to do it is that it's a good defense in a lawsuit. Like you say, all of this is a legal minefield.
So my understanding was that the original code was specifically not fed into Claude. But was almost certainly part of its training data, which complicates things, but if that's fair use then it's not relevant? If training's not fair use and taints the output, then new-chardet is a derivative of a lot of things, not just old-chardet...
This is all new legal ground. I'm not sure if anyone will go to court over chardet, though, but something that's an actual money-maker or an FSF flagship project like readline, on the other hand, well that's a lot more likely.
My understanding is they did do the dance. From the article: "He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch."
One could still make the argument that using the test suite was a critical contributing factor, but it is not a part of the resulting library. So in my uninformed opinion, it seems to me like the clean room argument does apply.
Why are people even having problems with sharing their changes to begin with? Just publishing it somewhere does not seem too expensive. The risk of accidentally including stuff that is not supposed to become public? Or are people regularly completely changing codebases and do not want to make the effort freely available, maybe especially to competitors? I would have assumed that the common case is adding a missing feature here, tweaking something there, if you turn the entire thing on its head, why not have your own alternative solution from scratch?
> an argument for protecting that test suite and API specification under copyleft terms.
If we protect API under copyright, it makes it easier to prevent interoperability. We obviously do NOT want that. It would give big companies even more power.
Now in the US, the Supreme Court that the output of an LLM is not copyrightable. So even a permissive licence doesn't work for that reimplementation: it should be public domain.
Disclaimer: I am all for copyleft for the code I write, but already without LLMs, one could rewrite a similar project and use the licence they please. LLMs make them faster at that, it's just a fact.
Now I wonder: say I vibe-code a library (so it's public domain in the US), I don't publish that code but I sell it to a customer. Can I prevent them from reselling it? I guess not, since it's public domain?
And as an employee writing code for a company. If I produce public domain code because it is written by an LLM, can I publish it, or can the company prevent me from doing it?
If you decide to improve it in any way to fit your needs you can merely tell your own AI to re-implement it with your changes. Then it's proprietary to you.
i've been following this for a while.. and the trend for copyright (of any form - books code pictures music whatever) being laundered by reinventing the "same" thing in-some-way.. is kind-of clear.
But what happens with the new things? Has the era of software-making (or creating things at large) finished, and from now on everything will be re-(gurgitated|implemented|polished) old stuff?
Or all goes back to proprietary everything.. Babylon-tower style, noone talks to noone?
edit: another view - is open-source from now on only for resume-building? "see-what-i've-built" style
Both sides are wrong on this actually. Computer generated code has no copyright protection.
>The U.S. Copyright Office (USCO) and federal courts have consistently ruled that AI-generated works—where the expressive elements are determined by the machine, even in response to a human prompt—lack the necessary human creative input and therefore cannot be copyrighted.
All this code is public domain. Your employees can publish "your" AI generated code freely and it won't matter how many tokens you spent generating it. It is not covered by copyright.
I feel like the licenses that suffer the most isn't the GPL, but the ones like SSPL. If your code can be re-implemented easily and legally by AWS using an LLM, why risk publishing it?
It does feel like open source is about to change. My hunch is that commercial open source (beyond the consultation model) risks disappearing. Though I'd be happy to be proven wrong.
I don't think this part is correct: "If you distribute modified code, or offer it as a networked service, you must make the source available under the same terms."
One of the things that irks me about this whole thing is, if it’s so clean room and distinct, why make the changes to the existing project? Why not make an entirely new library?
The answer to that, I think, is that the authors wanted to squat an existing successful project and gain a platform from it. Hence we have news cycle discussing it.
Nobody cares about a new library using AI, but squash an existing one with this stuff, and you get attention. It’s the reputation, the GitHub stars, whatever
I mean, Blanchard was the longtime maintainer of chardet already, and had wanted to relicense it for years. So I think that complicates your picture of "squatting an existing successful project".
Honestly it's a weird test case for this sort of thing. I don't think you'd see an equivalent in most open source projects.
See also "A Declaration of the Independence of Cyberspace" (https://www.eff.org/cyberspace-independence), and what a goofy, naive, misguided disaster that early internet optimism turned into.
No, AI does not mean the end of either copyright or copyleft, it means that the laws need to catch up. And they should, and they will.
1) Legality and morality are obviously different and unrelated concepts. More people should understand that.
2) Copyright was the wrong mechanism to use for code from the start, LLMs just exposed the issue. The thing to protect shouldn't be creativity, it should be human work - any kind of work.
The hard part of programming isn't creativity, it's making correct decisions. It's getting the information you need to make them. Figuring out and understanding the problem you're trying to solve, whether it's a complex mathematical problem or a customer's need. And then evaluating solutions until you find the right one. (One constrains being how much time you can spend on it.)
All that work is incredibly valuable but once the solution exists, it's each easier to copy without replicating or even understanding the thought process which led to it. But that thought process took time and effort.
The person who did the work deserved credit and compensation.
And he deserves it transitively, if his work is used to build other works - proportional to his contribution. The hard part is quantifying it, of course. But a lot of people these days benefit from throwing their hands up and saying we can't quantify it exactly so let's make it finders keepers. That's exploitation.
3) Both LLM training and inference are derivative works by any reasonable meaning of those words. If LLMs are not derivative works of the training data then why is so much training data needed? Why don't they just build AI from scratch? Because they can't. They just claim they found a legal loophole to exploit other people's work without consent.
I am still hoping the legal people take time to understand how LLMs work, how other algorithms, such as synonym replacement or c2rust work, decide that calling it "AI" doesn't magically remove copyright and the huge AI companies will be forced to destroy their existing models and train new ones which respect the licenses.
Imagine if the author has his way, and when we have AI write software, it becomes legally under the license of some other sufficiently similar piece of software. Which may or may not be proprietary. "I see you have generated a todo app very similar to Todoist. So they now own it." That does not seem like a good path either for open source software or for opening up the benefits of AI generated software.
What if someone doesn't declare that it has been reimplemented using an LLM? Isn't it enough to simply declare that you have reimplemented the software without using an LLM? Good luck proving that in court...
One thing is certain, however: copyleft licenses will disappear: If I can't control the redistribution of my code (through a GPL or similar license), I choose to develop it in closed source.
That's a non-sequitur. chardet v7 is GPL-derived work (currently in clear violation of the GPL). If xe wanted it to be a different thing xe should've published as such. Simple as.
If the model wasn't trained on copyleft, if he didn't use a copyleft test suite and if he wasn't the maintainer for years. Clearly the intent here is copyright infringement.
If you have software your testsuite should be your testsuite, you do dev with a testsuite and then mit without releasing one. Depending on the test-suite it may break clean room rules, especially for ttd codebases.
I think what is happening is the collapse of the “greater good”. Open source is dependent upon providing information for the greater good and general benefit of its readers. However now that no one is reading anything, its purpose is for the great good of the most clever or most convincing or richest harvester.
> Ronacher notes this as an irony and moves on. But the irony cuts deeper than he lets on. Next.js is MIT licensed. Cloudflare's vinext did not violate any license—it did exactly what Ronacher calls a contribution to the culture of openness, applied to a permissively licensed codebase. Vercel's reaction had nothing to do with license infringement; it was purely competitive and territorial. The implicit position is: reimplementing GPL software as MIT is a victory for sharing, but having our own MIT software reimplemented by a competitor is cause for outrage. This is what the claim that permissive licensing is “more share-friendly” than copyleft looks like in practice. The spirit of sharing, it turns out, runs in one direction only: outward from oneself.
This argument makes no sense. Are they arguing that because Vercel, specifically, had this attitude, this is an attitude necessitated by AI, reimplementation, and those who are in favor of it towards more permissive licenses? That certainly doesn't seem to be an accurate way to summarize what antirez or Ronacher believe. In fact, under the legal and ethical frameworks (respectively) that those two put forward, Vercel has no right to claim that position and no way to enforce it, so it seems very strange to me to even assert that this sort of thing would be the practical result of AI reimplementations. This seems to just be pointing towards the hypocrisy of one particular company, and assuming that this would be the inevitable universal, attitude, and result when there's no evidence to think so.
It's ironic, because antirez actually literally addresses this specific argument. They completely miss the fact that a lot of his blog post is not actually just about legal but also about ethical matters. Specifically, the idea he puts forward is that yes, corporations can do these kinds of rewrites now, but they always had the resources and manpower to do so anyway. What's different now is that individuals can do this kind of rewrites when they never have the ability to do so before, and the vector of such a rewrite can be from a permissive to copyleft or even from decompile the proprietary to permissive or copyleft. The fact that it hasn't been so far is a more a factor of the fact that most people really hate copyleft and find an annoying and it's been losing traction and developer mind share for decades, not that this tactic can't be used that way. I think that's actually one of the big points he's trying to make with his GNU comparison — not just that if it was legal for GNU to do it, then it's legal for you to do with AI, and not even just the fundamental libertarian ethical axiom (that I agree with for the most part) that it should remain legal to do such a rewrite in either direction because in terms of the fundamental axioms that we enforce with violence in our society, there should be a level playing field where we look at the action itself and not just whether we like or dislike the consequences, but specifically the fact that if GNU did it once with the ability to rewrite things, it can be done again, even in the same direction, it now even more easily using AI.
> They completely miss the fact that a lot of his blog post is not actually just about legal but also about ethical matters.
Honestly I was confused about the summarization of my blog post into just a legal matter as well. I hope my blog post will be able to flash at least a short time in the HN front page so that the actual arguments it contain will get a bit more exposure.
I'm failing to see what in the quoted text you took to be about AI rewrites specifically? It just reads as a slightly catty aside about the social reaction of rewrites in general (by implying the one example is generalizable.)
I think AI is very much eroding the legitimacy of copyright - at least to software, which is long been questioned since it's more like math than creative expression.
I think the industry will realize that it made a huge mistake by leaning on copyright for protection rather than on patents.
Probably a wiser approach is to consider different times require different measures (in general!).
I did not study in detail if copyright "has always been nonsense", but I do agree that nowadays some of the copyright regulations are nonsense (for example the very long duration of life + 70 years)
IMO the core idea of copyright isn't nonsense, but I do think the current implementation (70+ years after death) is egregiously overpowered. I've always thought the current laws were too deeply entrenched to ever change, but I'm tentatively optimistic AI will shock the system hard enough to trigger actual reform.
Actually I think the last 20 years of the Internet demonstrates that copyright is more important than ever, because unless it's enforced, people with more capital than the copyright owner will simply steal creative works and profit from them.
The idea that "information wants to be free" was always a lie, meant to transfer value from creators to platform owners. The result of that has been disastrous, and it's long past time to push the pendulum in the other direction.
I think we're going one step too far even, AI itself is a gray area and how can they guarantee it was trained legally or if it's even legal what they're doing and how can they assert that the input training data didn't contain any copyrighted data.
Google already spent billions of dollars and decades of lawyer hours proving it out as fair use. The legal challenges we see now are the dying convulsions of an already broken system of publishers and IP hoarders using every resource at their disposal to manipulate authors and creators and the public into thinking that there's any legitimacy or value underlying modern copyright law.
AI will destroy the current paradigm, completely and utterly, and there's nothing they can do to stop it. It's unclear if they can even slow it, and that's a good thing.
We will be forced to legislate a modern, digital oriented copyright system that's fair and compatible with AI. If producing any software becomes a matter of asking a machine to produce it - if things like AI native operating systems come about, where apps and media are generated on demand, with protocols as backbone, and each device is just generating its own scaffolding around the protocols - then nearly none of modern licensing, copyright, software patents, or IP conventions make any sense whatsoever.
You can't have horse and buggy traffic conventions for airplanes. We're moving in to a whole new paradigm, and maybe we can get legislation that actually benefits society and individuals, instead of propping up massive corporations and making lawyers rich.
Google has cut out some very specific ruling that have nothing to do with modern AI. These systems are just a really slow/lossy git clone, current law has no trouble with it, it's broadly illegal.
If corporations are allowed to launder someone else work as their own people will simply stop working and just start endlessly remixing a la popular music.
> If source code can now be generated from a specification, the specification is where the essential intellectual content of a GPL project resides.
Our foreparents fought for the right to implement works-a-like to corporate software packages, even if the so-called owners did not like it. We're ready to throw it all away, and let intellectual property owners get so much more control.
The implications will not end up being anti-large-corporation or pro-sharing. If you can prevent someone from re-implementing a spec or building a client that speaks your API or building a work-a-like, it will be the large corporations that exersize this power as usual.
Sounds very similar to that whole API lawsuit with oracle.
And the whole Adobe pdf thing and the whole Microsoft word thing. And the whole ibm pc thing. Imagine if we were forced to keep using ibm from when they lost their way until now simply because anti-ai luddites were able to scare monger
I would wager that the vast majority of people commenting here about the pitfalls of AI, especially as it relates to governance and laws, are heavy users of AI, recognize the import and value it brings, and find ways to utilize it more themselves, so not sure using an ad-hominem dismissal of very valid objections are going to be effective.
(side bar: the phrase "anti-<whatever> luddites" is way, way overused, especially here. Let's get more creative, people!)
>Our foreparents fought for the right to implement works-a-like to corporate software packages, even if the so-called owners did not like it
Our "foreparents" weren't competing with corporations with unlimited access to generative AI trained on their work. The times, they're-a-changin'.
You're rehashing the argument made in one of the articles which this piece criticizes and directly addresses, while ignoring the entirety of what was written before the conclusion that you quoted.
If anyone finds themselves agreeing with the comment I'm responding to, please, do yourself a favor and read the linked article.
I would do no justice to it by reiterating its points here.
I believe the GP post is saying that if we react to the new AI-enabled environment by arbitrarily strengthening IP controls for IP owners, the greatest benefactors will almost certainly be lawyer-laden corporations, not communities, artists, or open source projects. That seems like a reasonable argument.
It seems like the answer is to adjust IP owner rights very carefully, if that's possible. It sounds very hard, though.
I mean. Yeah. GPL's genius was that it used Copyright, which proprietary enterprise wouldn't dare dismantle, to secure for the public a permanent public good.
Pretty sure no one, (but me anyway) saw overt theft of IP by ignoring IP law through redefinition coming. Admittedly I couldn't articulate for you capital would skill transfer and commoditize it in the form of pay to play data centers, but give me a break, I was a teenager/twenty something at the time.
Would software be more or less free in a world without copyright?
I argue more free. EULAs and restrictions on how+for what software can be used, like DRM, typically use copyright as their legal backing. GPL licenses turn that on it's head but that doesn't redeem the original, flawed, law.
This seems to follow the letter but not the spirit of the license. If this does pass legal muster, we can do the same to whatever proprietary software we wish, which makes a dramatically different but IMO better ecosystem in the end.
The really interesting question to me is if this transcends copyright and unravels the whole concept of intellectual property. Because all of it is premised on an assumption that creativity is "hard". But LLMs are not just writing software, they are rapidly being engineered to operate completely generally as knowledge creation engines: solving math proofs, designing drugs, etc.
So: once it's not "hard" any more, does IP even make sense at all? Why grant monopoly rights to something that required little to no investment in the first place? Even with vestigial IP law - let's say, patents: it just becomes and input parameter that the AI needs to work around the patents like any other constraints.
> So: once it's not "hard" any more, does IP even make sense at all? Why grant monopoly rights to something that required little to no investment in the first place? Even with vestigial IP law - let's say, patents: it just becomes and input parameter that the AI needs to work around the patents like any other constraints.
I think it still does: IIRC, the current legal situation is AI-output does not qualify for IP protections (at least not without substantial later human modification). IP protections are solely reserved for human work.
And I'm fine with that: if a person put in the work, they should have protections so their stuff can't be ripped off for free by all the wealthy major corporations that find some use for it. Otherwise: who cares about the LLMs.
What if a person puts in the work, but the work was worthless or can be trivially reproduced without effort?
See also: https://en.wikipedia.org/wiki/Sweat_of_the_brow
I think you have a rather idealized model of IP in mind. In practice, IP law tends to be an expensive weapon the wealthy major corporations use against the little guy. Deep enough pockets and a big enough warchest of broad parents will drain the little guy every time.
> In practice, IP law tends to be an expensive weapon the wealthy major corporations use against the little guy. Deep enough pockets and a big enough warchest of broad parents will drain the little guy every time.
Then fix that instead of blowing it up. Because IP law is also literally the only thing that protects the little guy's work in many cases.
Arguments like yours are kinda unfathomably incomplete to me, almost like they're the remnants of some propaganda campaign. It's constructed to appeal to the defense of the little guy, but the actual effect would be to disempower him and further empower the wealthy major corporations with "big enough warchest[s]."
I mean, one thing I think the RIAA would love is to stop paying royalties to every artist ever. And the only thing they'd be worried about is an even bigger fish (like Amazon, Apple, or Spotify) no longer paying royalties to them. But as you said, they have a big enough war chest that they probably could force a deal somehow. All the artists without a war chest? Left out in the cold.
Blowing up IP would sink the RIAA. They would no longer have legal grounds to go after file sharing, and I’m confident that given the same legal footing that file sharing would win any day of the week.
> AI-output does not qualify for IP protections
I beg to differ. AI-output did not entitle the person creating the prompt for IP protections, so far – but my objection is not directed towards the "so far", but towards your omission of "the person creating the prompt", because if an AI outputs copyrighted material from the training data, that material is still copyrighted. AI is not a magical copyright removal machine.
The U.S. Supreme Court just declined to hear a case, thus upholding a lower court precedent that LLM output are not copyrightable: https://www.reuters.com/legal/government/us-supreme-court-de...
What this means in practice is that (currently), all output of an LLM is legally considered to not be copyrightable (to the extent that it's an original work). If it happens to regurgitate an existing copyrighted work, though, is that infringement? I'm not sure we have a legal precedent on that question yet.
There’s several large settlements that say Anthropomorphic/OAI didn’t want to have legal precedent. In general if it’s not outright regurgitated it would be derivative.
Does this matter in practice though? By modifying some of the generated code and not taking a solution produced by an LLM end-to-end but borrowing heavily from it, can't a human claim full ownership of the IP even though in reality the LLM did most of the relevant work?
I think as long as the human puts in substantial and transformational effort, they can claim to be the copyright holder of the entire work, yes.
Good. Intellectual property is now a twisted concept by the elite, whatever its benefits were previously. As soon as Disney made Mickey popular, it was all downhill.
More likely: this is a transitional phase where our previously hard problems become easy, and we will soon set our sights on new and much harder problems. The pinnacle of creative achievement in the universe is probably not 2010s B2B SaaS.
It is entirely possible, however, that human beings will not be the primary drivers of progress on those problems.
The basis of your argument is that AI-generated work isn't hard, but your conclusion is that ALL work, AI-generated or not, should lose IP rights?
It might unravel intellectual property, just not in a fair way. When capitalism started, public land was enclosed to create private property. Despite this being in many cases a quite unfair process, we still respect this arrangement.
With AI, a similar process is happening - publicly available information becomes enclosed by the model owners. We will probably get a "vestigial" intellectual property in the form of model ownership, and everyone will pay a rent to use it. In fact, companies might start to gatekeep all the information to only their own LLM flavor, which you will be required to use to get to the information. For example, product documentation and datasheets will be only available by talking to their AI.
Nothing changes for drug patents regardless of whether an LLM was used in the discovery process.
Even if all I have to do is tell my agent, "here is a patent for a drug, analyse the patent and determine an equivalent but non-infringing drug" and it chugs away for a couple of hours and spits out a drug along with all the specifications to manufacture it?
I guess the state of play will be that for new drugs the original manufacturer will already have done that and ensured that literally anything that could be found as a workaround is included in the scope of the patent. But I feel like it will not be possible to keep that wartertight.
Yes, even so. Human drug researchers have been doing the same thing for decades. As soon as one pharmaceutical company launches a successful small-molecule drug everyone else jumps to find a minor tweak that will hit the same target (ideally with fewer side effects) while evading the patent. There is already specialized software to help with this process so I'm skeptical that LLM agents would be very helpful for this use case.
The formula is what is patented, not the process to come up with it.
Not sure why this should be true; the US Supreme Court recently chose to let precedent stand that AI creations are not copyrightable. https://www.reuters.com/legal/government/us-supreme-court-de...
That also seems relevant for this whole discussion, actually -- if a work can't be copyrighted it certainly can't have a changed license, or any license at all. (I guess it's effectively public domain to the extent that it's public at all?)
You're really missing the point in multiple ways. First, precedents on copyright law are irrelevant to patent law. Second, AI generated works generally can be copyrighted under the human creator's name.
No, I think you are quite incorrect, at least on the latter point:
"Lower courts upheld a U.S. Copyright Office decision that the AI-crafted visual art at issue in the case was ineligible for copyright protection because it did not have a human creator."
Not eligible for copyright protection does not mean it can be copyrighted "under the human creator's name". It means there is no creative work at all. No copyright.
And while courts in theory aren't supposed to apply copyright precedent to patent cases, in practice, they apparently do a lot of the time, so it's kind of a mess! https://scholarship.kentlaw.iit.edu/ckjip/vol16/iss1/4/#:~:t...
If you think about creative outcomes as n dimensional 'volumes', AI expressions can cover more than humans in many domains. These are precisely artistic styles, music styles etc. and tbh not everyone can be a Mozart but may be a lot more with AI can be Mozart lite. This begs the question how much of creativity is appreciated as a shared experience
"Hard" or "easy" has never been part of the premise.
A company spends a decade and billions of dollars to develop a groundbreaking drug and patents it.
I think of a cool new character called "Mr Poop" and publish a short story about him with an hour of work.
Both of us get the exact same protection under the law (yes yes I know copyright vs patent etc., but ultimately they are all about IP protection).
I've always thought the opposite: IP law was created to make sure creativity stays hard, and hence controllable by the elites.
Patents came along when farmers started making city goods, threatening guilds secrets. Copyright came when the printing press made copying and translating the bible easy and accessible to all. (Trademark admittedly does not fit this view, but doesn't seem all that damaging either)
To Protect The Arts, and To Time Limit Trade Secrets were just the Protect The Children of old times, a way to confuse people who didn't look too hard at actual consequences.
This means that the future of IP depends on what lets the powers that be pull up the ladder behind them. Long term I'd expect e.g. copyright expansion and harder enforcement, just because cloning by AI gets easy enough to threaten the status quo.
:/ before copyright you just had patrons, which looks a lot more like the rich controlling what art gets made than what we have today
> Trademark admittedly does not fit this view, but doesn't seem all that damaging either
Isn’t trademark the only thing keeping a certain cartoon mouse out of the public domain, despite the fact that his earliest animations are out of copyright? Not sure if you’d consider that damaging, or if anyone has yet tested the boundaries of the House of Mouse’s patience here.
Don't worry. The courts have consistently sided with huge companies on copyright. In the US. In Europe. Doesn't matter.
Company incorporates GPL code in their product? Never once have courts decided to uphold copyright. HP did that many times. Microsoft got caught doing it. And yet the GPL was never applied to their products. Every time there was an excuse. An inconsistent excuse.
Schoolkid downloads a movie? 30,000 USD per infraction PLUS armed police officer goes in and enforces removal of any movies.
Or take the very subject here. AI training WAS NOT considered fair use when OpenAI violated copyright to train. Same with Anthropic, Google, Microsoft, ... They incorporated harry potter and the linux kernel in ChatGPT, in the model itself. Undeniable. Literally. So even if you accept that it's changed now, OpenAI should still be forced to redistribute the training set, code, and everything needed to run the model for everything they did up to 2020. Needless to say ... courts refused to apply that.
So just apply "the law", right. Courts' judgement of using AI to "remove GPL"? Approved. Using AI to "make the next Disney-style movie"? SEND IN THE ARMY! Whether one or the other violates the law according to rational people? Whatever excuse to avoid that discussion is good enough.
Someone should put this to the test. Take the recently leaked Minecraft source code and have Copilot build an exact replica in another programming language and then publish it as open source. See if Microsoft believes AI is copyright infringement or not.
As described, this would not be the same thing. If the AI is looking at the source and effectively porting it, that is likely infringement. The idea instead should be "implement Minecraft from scratch" but with behavior, graphics, etc. identical. Note that you'll need to have an AI generate assets or something since you can't just reuse textures and models.
AI models have already looked at the source of GPL software and contain it in their dataset. Adding the minecraft source to the mix wouldn't seem much different. Of course art assets and trade marks would have to be replaced. But an AI "clean room" implementation has yet to be legally tested.
That's why he is saying it's not equivalent. For it to be the same, the LLM would have to train on/transform Minecraft's source code into its weights, then you prompt the LLM to make a game using the specifications of Minecraft solely through prompts. Of course it's copyright infringement if you just give a tool Minecraft's source code and tell it to copy it, just like it would be copyright infringement if you used a copier to copy Minecraft's source code into a new document and say you recreated Minecraft.
What if Copilot was already trained with Minecraft code in the dataset? Should be possible to test by telling the model to continue a snippet from the leaked code, the same way a news website proved their articles were used for training.
Is there a legal distinction between training, post-training, fine tuning and filling up a context window?
In all of these cases an AI model is taking a copyrighted source, reading it, jumbling the bytes and storing it in its memory as vectors.
Later a query reads these vectors and outputs them in a form which may or may not be similar to the original.
Judges have previously ruled that training counts as sufficiently transformative to qualify for fair use: https://www.whitecase.com/insight-alert/two-california-distr...
I don't know of any rulings on the context window, but it's certainly possible judges would rule that would not qualify as transformative.
For copyright purposes I think there is an important legal distinction between training data (fed in once, ahead of time, and can in theory no longer be recovered as-is) and context window data (stored exactly for the duration of the model call).
I'm not sure there should be, but I think there is.
A room "as clean" as the one under dispute (chardet) is very easy to replicate.
AI 1: - (reads the source), creates a spec + acceptance criteria
AI 2: - implements from spec
AI 1 is in the position of the maintainer who facilitated the license swap.
this is the question of the hour. Imagine using this LLM proxy to license-strip major parts of leaked Windows source code to produce code for WINE.
On top of all of this, there are the attempts at binary decompilation using LLMs and other new tools that have been discussed on this site recently.
I’ve often thought that the key to fighting this is through this exact method. Turn the tool against them
The big question is: if copyrighted material was used in the training material, is the LLM's output copyright infringement when it resembles the training material? In your example, you are taking the copyrighted material and giving it to the LLM as input and instructing the LLM to process it. Regardless of where the legal cards fall, this is a much less ambiguous scenario.
I think Disney ran into this with people generating Marvel characters etc
Fictional characters can have their own copyright.
https://www.nolo.com/legal-encyclopedia/protecting-fictional...
https://en.wikipedia.org/wiki/Copyright_protection_for_ficti...
This was not about legality.
> That question is this: does legal mean legitimate?
Just because something is legal does not mean it's moral thing to do.
this question should've been posed earlier when first LLMs were training. many people chose to ignore the question, and now, several distillation epochs later, it is not a question that matters, as both yes/no are true, and not true.
is it legitimate for millions of people to exploit, expound on knowledge that was perhaps, to begin with, not legitimate to use? well they did already, who's to judge the commons now?
What a ridiculous take. Many people loudly raised the question and objected to the practice from the beginning, but a handful of companies ignored the objections and ran faster than the legal system. If they were in the wrong, legally or morally, they still deserve to face repercussions for it.
You will probably run into design patents.
Software patents is not a thing in EU.
I said design patents:
https://en.wikipedia.org/wiki/Design_patent
Very much is. "Software programs, as such" are exempt in the EPC article 52. However if the software program interacts with the world - if it has a "further technical effect" - it is patentable.
https://en.wikipedia.org/wiki/Software_patents_under_the_Eur...
They absolutely are. That's a myth.
But also software patents and design patents are totally different things.
They might not care. Products win not by quality or features but by advertisement, hype and network effects.
The original implementation would still have the upper hand here. OTOH if I as a nobody create something cool, there's nothing stopping a huge corporation from "reimplementing" (=stealing) it and and using their huge advertising budget to completely overshadow me.
And that's how they like it.
Given how hard companies like Nintendo and Microsoft have been taking down leaks or fan creations, it seems they very much do care about keeping this stuff locked down.
I believe it is a narrow view of the situation. If we take a look into the history, into the reasons for inventing GPL, we'll see that it was an attempt to fight copyrights with copyrights. The very name 'copyleft' is trying to convey the idea.
What AI are eroding is copyright. You can re-implement not just a GPL program, but to reverse engineer and re-implement a closed source program too, people have demonstrated it already, there were stories here on HN about it.
AI is eroding copyright, so there may no longer be a need for the GPL. GNU should stop and rethink its stance, chuck away the GPL as the main tool to fight evil software corporations and embrace LLM as the main weapon.
> LLM as the main weapon
LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source where you could do innovative work on your own computer running Linux or FreeBSD or some other open OS.
I don't think that's an exciting idea for the Free Software Foundation.
Perhaps with time we'll be able to run local ones that are 'good enough', but we're not there yet.
There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
Edit: I guess the conclusion I come to is that LLM's are good for 'getting things done', but the context in which they are operating is one where the balance of power is heavily tilted towards capital, and open source is perhaps less interesting to participate in if the machines are just going to slurp it up and people don't have to respect the license or even acknowledge your work.
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source
Yeah, a bit of a conundrum. But I don't think that fighting for copyright now can bring any benefits for FOSS. GNU should bring Stallman back and see whether he can come with any new ideas and a new strategy. Alternatively they could try without Stallman. But the point is: they should stop and think again. Maybe they will find a way forward, maybe they won't but it means that either they could continue their fight for a freedom meaningfully, or they could just stop fighting and find some other things to do. Both options are better then fighting for copyright.
> There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
I want a clarify this statement a bit. The thing with LLM relying on work of others are not against GPU philosophy as I understand it: algorithms have to be free. Nothing wrong with training LLMs on them or on programs implementing them. Nothing wrong with using these LLMs to write new (free) programs. What is wrong are corporations reaping all the benefits now and locking down new algorithms later.
I think it is important, because copyright is deemed to be an ethical thing by many (I think for most people it is just a deduction: abiding the law is ethical, therefore copyright is ethical), but not for GNU.
>Yeah, a bit of a conundrum.
IMO the primary significant trend in AI. Doesn't get talked about nearly enough. Means the AI is working, I guess.
>GNU should bring Stallman back ... Alternatively they could try without Stallman.
Leave Britney alone >:(
>copyright is deemed to be an ethical thing by many (I think for most people it is just a deduction: abiding the law is ethical, therefore copyright is ethical)
I've busted out "intellectual property is a crime against humanity" at layfolk to see if that shortcuts through that entire little politico-philosophical minefield. They emote the requisite mild shock when such things as crimes against humanity are mentioned; as well as at someone making such a radical statement which seems to come from no familiar species of echo chamber; and then a moment later they begin to very much look like they see where I'm coming from.
How do you even argue such a thing? I've had no such luck, I've met many people who seem to view copyright and a person owning their ideas and work as a sort of inherent moral.
Not saying this gets through to people, but copyright is purely about the legal ability to restrict what other people do. Whereas property rights are about not allowing others to restrict what you do (e.g. by taking your stuff).
>Perhaps with time we'll be able to run local ones that are 'good enough', but we're not there yet.
Right now, we can get local models that you can run on consumer hardware, that match capabilities of state of the art models from two years ago. The improvements to model architecture may or may not maintain the same pace in the future, but we will get a local equivalent to Opus 4.6 or whatever other benchmark of "good enough" you have, in the foreseeable future.
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones
There are near-SOTA LLM's available under permissive licenses. Even running them doesn't require prohibitive expenses on hardware unless you insist on realtime use.
> running them doesn't require prohibitive expenses on hardware
What async tasks could a local LLM accomplish on Intel 11th gen CPU with 32GB RAM?
> LLM's - to date - seem to require massive capital expenditures to have the highest quality ones, which is a monumental shift in power towards mega corporations and away from the world of open source where you could do innovative work on your own computer running Linux or FreeBSD or some other open OS.
When the FSF and GPL were created, I don't think this was really a consideration. They were perfectly happy with requiring Big Iron Unix or an esoteric Lisp Machine to use the software - they just wanted to have the ability to customize and distribute fixes and enhancements to it.
Maybe a good open source idea is to "seti at home" style crowd-source training, assuming that's possible.
> There's also an ethical/moral question that these things have been trained on millions of hours of people's volunteer work and the benefits of that are going to accrue to the mega corporations.
This was already the case and it just got worse, not better.
At a certain point, I think we had reached a kind of equilibrium where some corporations were decent open source citizens. They understood that they could open source things like infrastructure or libraries and keep their 'crown jewels' closed. And while Stallman types might not have been happy with that, it seemed to work out for people.
Now they've just hoovered up all the free stuff into machines that can mix it up enough to spit it out in a way that doesn't even require attribution, and you have to pay to use their machine.
AI essentially gatekeeps all of open source to companies to pluck from to their hearts content. And individual contributors using these tools and freely mixing it with their own - usual minor - contributions are another step of whitewashing because they're definitely not going to own up to writing only 5% of the stuff they got paid for.
Before we had RedHat and Ubuntu, who at least were contributing back, now we have Microsoft, Anthropic and OpenAI who are racing to lock the barn door around their new captive sheep. It's just a massive IP laundromat.
Is massive capital expenditure not also required to enforce the GPL? If some company steals your GPLed code and doesn't follow the license, you will have to sue them and somebody will have to pay the lawyers.
> Is massive capital expenditure not also required to enforce the GPL?
It's nowhere near the order of magnitude of the kind of spending they're sinking into LLM's. The FSF and other groups were reasonably successful at enforcing the GPL, operating on a budget 1000's of times smaller than that of AI companies.
Right but LLM companies are building frontier models with frontier talent while trying to sock up demand with a loss leader strategy, on top of an historic infrastructure build out.
Being able to coat efficiently run frontier models is i think, not a high priced endeavor for an org (compared to an individual).
IMO the proposition is little fishy, but its not totally without merit and imo deserves investigation. If we are all worried about our jobs, even via building custom for sale software, there is likely something there that may obviate the need at least for end user applications. Again, im deeply skeptical, but it is interesting.
> Being able to coat efficiently run frontier models is i think, not a high priced endeavor for an org
Running proprietary model would make you subject to whatever ToS the LLM companies choose on a particular day, and what you can produce with them, which circles back to the raison d'etre for the GPL and GNU.
Until all software copyright is dead and buried, there is no need for copyleft to change tack. Otherwise there rising tide may rise high enough to drown GPL, but not proprietary software.
Open source is easier to counterfeit/license-launder/re-implement using LLMs because source code is much lower-hanging fruit, and is understood by more people than closed-source assembly.
How close are we to good enough and who's working on that? I would be interested in supporting that work; to my mind, many of the real objections to LLMs are diminished if we can make them small and cheap enough to run in the home (and, perhaps, trained with distributed shared resources, although the training problem is the harder one).
Good question. It seems like most of the tech world is perfectly happy to be sharecroppers on the Big AI farms. I guess that's not quite the right analogy, since they're doing their own things with it; just that at the end of the day, the tool they're building everything on is owned by someone else.
Copyleft is a mirror of copyright, not a way to fight copyright. It grants rights to the consumer where copyright grants rights to the creator. Importantly, it gives the end-user the right to modify the software running on their devices.
Unfortunately, there are cases where you simply can't just "re-implement" something. E.g., because doing so requires access to restricted tools, keys, or proprietary specifications.
These are words of Stallman:
"So, I looked for a way to stop that from happening. The method I came up with is called “copyleft.” It's called copyleft because it's sort of like taking copyright and flipping it over. [Laughter] Legally, copyleft works based on copyright. We use the existing copyright law, but we use it to achieve a very different goal."
https://writings.hongminhee.org/2026/03/legal-vs-legitimate/
> flipping it over.
i.e. mirroring it
> use it to achieve a very different goal."
"very different goal" isn't the same as "fundamentally destroying copyright"
the very different goal include to protect public code to stay public, be properly attributed, prevent companies from just "sizing" , motivate other to make their code public too etc.
and even if his goals where not like that, it wouldn't make a difference as this is what many people try to archive with using such licenses
this kind of AI usage is very much not in line with this goals,
and in general way cheaper to do software cloning isn't sufficient to fix many of the issues the FOSS movement tried to fix, especially not when looking at the current ecosystem most people are interacting with (i.e. Phones)
---
("sizing"): As in the typical MS embrace, extend and extinguish strategy of first embracing the code then giving it proprietary but available extensions/changes/bug fixes/security patches to then make them no longer available if you don't pay them/play by their rules.
---
Through in the end using AI as a "fancy complicated" photocopier for code is as much removing copyright as using a photocopier for code would. It doesn't matter if you use the photocopier blind folded and never looked at the thing you copied.
> We use the existing copyright law, but we use it to achieve a very different goal.
For the right goal, he should have called it "rightcopy".
That’s not a rebuttal of the OP’s point. None of that says anything about fighting copyright. It literally says he flipped it which is wha the OP said when they said it’s a mirror.
> It grants rights to the consumer where copyright grants rights to the creator.
It also grants one major right/feature to the creator, the ability to spread their work while keeping it as open as they intend.
> AI is eroding copyright, so there may no longer be a need for the GPL. GNU should stop and rethink its stance, chuck away the GPL as the main tool to fight evil software corporations and embrace LLM as the main weapon.
Is this LLM thing freely available or is it owned and controlled by these companies? Are we going to rent the tools to fight "evil software corporations"?
There already are LLMs with open weights that are better at code than state of the art closed source models from a year ago. For now, you most people may have to rent the hardware to run those models, since it's too expensive for most people to own something that can run inference on one trillion parameters, but I wouldn't consider LLMs to be controlled by "evil software corporations" at this point.
Open models do exist. They’re nowhere near aa good as frontier models, but they’re getting better all the time.
It’s probably only a matter of time before open models are as good as Claude code is today.
With the release of GLM-5, I would say that they are pretty much almost as good. Basically 90% as good as Opus 4.6 on most tasks for 20% of inference cost, and open weights.
easy, we ask Claude to write an open-source freely-available version of Claude with equal or better capabilities.
> chuck away the GPL as the main tool to fight evil software corporations and embrace LLM as the main weapon.
LLMs are one of the primary manifestations of 'evil software corporations' currently.
Until there is a capable open source open weight AI that is easily hostable by an average person - no, we still have a long way to go. You aren't going to have software freedom when the tool that enables it is controlled by a handful of powerful tech companies.
> we'll see that it was an attempt to fight copyrights with copyrights
it's not that simple
yes, GPLs origins have the idea of "everyone should be able to use"
but it also is about attribution the original author
and making sure people can't just de-facto "size public goods"
the kind of AI usage is removing attribution and is often sizing public goods in a way far worse then most companies which just ignored the license did
so today there is more need then ever in the last few decades for GPL like licenses
You've said "size" twice in comments, did you mean "seize"?
Its purpose "if you run the software you should be able to inspect and modify that software, and to share those modifications with your peers" not explicitly resist copyright. Yes copyright is bad in that it often prevents one from doing that, but it is not the purpose of the GPL to dismantle copyright.
Reducing it to "well you can clone the proprietary software you're forced to use by LLM" is really missing the soul of the GPL.
If not for copyright, you could always do that and copyleft wouldn't be needed.
Just because something is copyleft doesn't mean the person who gave you the binary you're using has to supply you with the code the used to build it. That's what the GPL does.
I agree with almost all of that, except the part about GNU changing their stance. I think GNU should stay true and consistent, if for no other reason than to not make many of their supporters who aren't on board with AI feel betrayed and have GNUs legacy soured. If the cause of LLMs conquering proprietary software needs an organization to champion it, let that be a new organization, not GNU.
That's naive. Copyright doesn't just apply to software. There already have been countless lawsuits about copying music long before the term "open source" was invented. No, changing the lyrics a bit doesn't circumvent copyright. Nor does translating a Stephen King novel to German and switching the names of the places and characters.
A court ordered the first Nosferatu movie to be destroyed because it had too many similarities to Dracula. Despite the fact that the movie makes rather large deviations from the original.
If Claude was indeed asked to reimplement the existing codebase, just in Rust and a bit optimized, that could well be a copyright violation. Just like rephrasing A Song ot Ice and Fire a bit, and switching to a different language, doesn't remove its copyright.
Claude was asked to implement a public API, not an entire codebase. The definition of a public API is largely functional; even in an unusually complex case like the Java standard facilities (which are unusually creative even in the structure and organization of the API itself) the reimplementation by Google was found to be fair use.
> Claude was asked to implement a public API, not an entire codebase.
Allegedly. There have been several people who doubted this story. So how to find out who is right? Well, just let Claude compare the sources. Coincidentally, Claude Opus 4.6 doesn't just score 75.6% on SWE-bench Verified but also 90.2% on BigLaw Bench.
It's like our copyright lawyer is conveniently also a developer. And possibly identical to the AI that carried out the rewrite/reimplemention in question in the first place.
> Just like rephrasing A Song ot Ice and Fire a bit, and switching to a different language, doesn't remove its copyright.
There is some precedent for this, e.g. Alchemised is a recent best seller that had just enough changed from its Harry Potter fan fiction source in order to avoid copyright infringement: https://en.wikipedia.org/wiki/Alchemised
(I avoided the term “remove copyright” here because the new work is still under copyright, just not Harry Potter - related copyright.)
That's apparently a different story with different plot, so that's not comparable.
Plots are broadly not copyrightable, “different plot” is less important than “different characters”.
This is naive. Advertisement and network effects win. Individuals cannot compete with corporations on equal ground here.
> What AI are eroding is copyright.
At the moment it's people that are eroding copyright. E.g. in this case someone did something.
"AI" didn't have a brain, woke up and suddenly decided to do it.
Realistically nothing to do with AI. Having a gun doesn't mean you randomly shoot.
> AI is eroding copyright
Unless it is IP of the same big corpos that consumed all content available. Good luck with eroding them.
So not only are we moving goalposts here, but we've decided the GNU team should join the other team? I don't understand how GNU would see mass model LLM training as anything but the most flagrant violations of their ethos. LLM labs, in their view, would be among the most evil software corporations to have ever existed.
While I personally agree with you, Richard Stallman (the creator of the GPL) does not. He has always advocated in favor of strong copyright protection, because the foundation of the GPL is the monopoly power granted by copyright. The problem that the GPL is intended to solve is proprietary software.
Generative models (AI) are not really eroding copyright. They are calling its bluff. The very notion of intellectual property depends on a property line: some arbitrary boundary where the property begins and ends. Generative models blur that line, making it impractical to distinguish which property belongs to whom.
Ironically, these models are made by giant monopolistic corporations whose wealth is quite literally a market valuation (stock price) of their copyrights! If generative models ever become good enough to reimplement CUDA, what value will NVIDIA have left?
The reality is that generative models are nowhere near good enough to actually call the bluff. Copyright is still the winning hand, and that is likely to continue, particularly while IP holders are the primary authors of law.
---
This whole situation is missing the forest for the trees. Intellectual Property is bullshit. A system predicated on monopoly power can only result in consolidated wealth driving the consolidation of power; which is precisely what has happened. The words "starving artist" ring every bit as familiar today as any time in history. Copyright has utterly failed the very goals it was explicitly written with.
It isn't the GPL that needs changing. So long as a system of copyright rules the land, copyleft is the best way to participate. What we really need is a cohesive political movement against monopoly power; one that isn't conveniently ignorant of copyright as its most significant source.
Right, anything that can be copied instantly for free cannot be realistically owned.
I think the missing thing here is that the license violation already happened. Most of the big models trained on data in a manner that violated terms of service. We'll need a court case but I think it's extremely reasonable to consider any model trained on GPL code to be infected with open licensing requirements.
You might wish that were true, but there are very strong arguments it's not. Training on copyleft licensed code is not a license violation. Any more than a person reading it is. In copyright terms, it's such an extreme transformative use that copyright no longer applies. It's fair use.
But agreed that we're waiting for a court case to confirm that. Although really, the main questions for any court cases are not going to be around the principle of fair use itself or whether training is transformative enough (it obviously is), but rather on the specifics:
1) Was any copyrighted material acquired legally (not applicable here), and
2) Is the LLM always providing a unique expression (e.g. not regurgitating books or libraries verbatim)
And in this particular case, they confirmed that the new implementation is 98.7% unique.
Transformative is not the only component of determining fair use, there’s also the economic displacement aspect. If you’re doing a book report and include portions of the original (or provide an interface for viewing portions à la Google Books) you aren’t a threat to the original authors ability to make a living.
If you’ve used copyrighted books and turned them into a free write-a-book machine, you are suddenly using the authors own works against them, in a way that a judge might rule is not very fair.
“ Effect of the use upon the potential market for or value of the copyrighted work: Here, courts review whether, and to what extent, the unlicensed use harms the existing or future market for the copyright owner’s original work. In assessing this factor, courts consider whether the use is hurting the current market for the original work (for example, by displacing sales of the original) and/or whether the use could cause substantial harm if it were to become widespread.”
https://www.copyright.gov/fair-use/
A human reading a unit of work is not a “copy”. I’m pretty sure our legal systems agree that thought or sight is not copying something.
Training an LLM inherently requires making a copy of the work. Even the initial act of loading it from the internet and copying it into memory to then train the LLM is a copy that can be governed by its license and copyright law
I think you are confusing two different meanings of the word ‘copy’. The fact that a computer loads it into memory does not make it automatically a ‘copy’ in the copyright sense.
> Training an LLM inherently requires making a copy of the work.
But that's not relevant here. Because the copyleft license does not prohibit that (and it's not even clear that any license can prohibit it, as courts may confirm it's fair use, as most people are currently assuming). That's why I noted under (1) that it's not applicable here.
It's absolutely prohibited to copy and redistribute for commercial purposes materials that you're unlicensed to do so with. This isn't an issue when it comes to the copy-left scenario (though it may potentially enforce transitive licensing requirements on the copier that LLM runners don't want to follow) but it is a huge issue that has come up with LLM training.
LLM training involves ingesting works (in a potentially transformative process) and partially reproduce them - that's a generally restricted action when it comes to licensing.
> It's absolutely prohibited to copy and redistribute for commercial purposes materials that you're unlicensed to do so with.
Sure, but that's not what LLM's generally do, and it's certainly not what they're intended to do.
The LLM companies, and many other people, argue that training falls under fair use. One element of fair use is whether the purpose/character is sufficiently transformative, and transforming texts into weights without even a remote 1-1 correspondence is the transformation.
And this is why LLM companies ensure that partial reproduction doesn't happen during LLM usage, using a kind of copyrighted-text filter as a last check in case anything would unintentionally get through. (And it doesn't even tend to occur in the first place, except when the LLM is trained on a bunch of copies of the same text.)
Yea, at the end of the day a big part of this question comes down to whether that copying is fair use and that is an open question with the transformative nature being the primary point in favor of the LLM. But it is copying from some works to another - if it doesn't have some fair use exception it is absolutely violating the licensing of most of the training data. It's a bit different from previous settled case law because it's copying so little from so many billions of different things. I think blocking reproduction is wise by LLM companies for PR purposes but it doesn't guarantee that training is a license exempted activity.
Yup. Of course it's copying. But all expectations are that courts will rule that fair use allows such copying, because of the nature of the transformation.
> Training on copyleft licensed code is not a license violation. Any more than a person reading it is.
Some might hold that we've granted persons certain exemptions, on account of them being persons. We do not have to grant machines the same.
> In copyright terms, it's such an extreme transformative use that copyright no longer applies.
Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim? Sure, it can also produce extremely transformed versions, but is that really relevant if it holds within it enough information for a (near-)verbatim reproduction?
>Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim? Sure, it can also produce extremely transformed versions, but is that really relevant if it holds within it enough information for a (near-)verbatim reproduction?
I feel as though, from an information-theoretic standpoint, it can't be possible that an LLM (which is almost certainly <1 TB big) can contain any substantial verbatim portion of its training corpus, which includes audio, images, and videos.
> We do not have to grant machines the same.
No we don't have to, but so far we do, because that's the most legally consistent. If you want to change that, you're going to need to pass new laws that may wind up radically redefining intellectual property.
> Has the model really performed an extreme transformation if it is able to produce the training data near-verbatim?
Of course it has, if the transformation is extreme, as it appears to be here. If I memorize the lyrics to a bunch of love songs, and then write my own love song where every line is new, nobody's going to successfully sue me just because I can sing a bunch of other songs from memory.
Also, it's not even remotely clear that the LLM can produce the training data near-verbatim. Generally it can't, unless it's something that it's been trained on with high levels of repetition.
I want to briefly pick at this:
> you're going to need to pass new laws that may wind up radically redefining intellectual property
You're correct that this is one route to resolving the situation, but I think it's reasonable to lean more strongly into the original intent of intellectual property laws to defend creative works as a manner to sustain yourself that would draw a pretty clear distinction between human creativity and reuse and LLMs.
> into the original intent of intellectual property laws to defend creative works as a manner to sustain yourself
But you're missing the other half of copyright law, which is the original intent to promote the public good.
That's why fair use exists, for the public good. And that's why the main legal argument behind LLM training is fair use -- that the resulting product doesn't compete directly with the originals, and is in the public good.
In other words, if you write an autobiography, you're not losing significant sales because people are asking an LLM about your life.
The big difference between people reading code and LLMs reading code is that people have legal liability and LLMs do not. You can't sue an LLM for copyright infringement, and it's almost impossible for users to tell when it happens.
BTW in 2023 I watched ChatGPT spit out hundreds of lines of F# verbatim from my own GitHub. A lot of people had this experience with GitHub Copilot. "98.7% unique" is still a lot of infringement.
> people have legal liability and LLMs do not. You can't sue an LLM for copyright infringement
That's not relevant, because you can still sue the person using the LLM and publishing the repository. Legal liability is completely unchanged.
>Legal liability is completely unchanged.
It's changed completely, from your own example.
If you comission art from an artist who paints a modified copy of Warhol's work, the artist is liable (even if you keep that work private, for personal use).
If you commission it from OpenAI (by sending a query to their ChatGPT API), by your argument, you are the person liable — and OpenAI is off the hook even if that work is distributed further.
I'm not going to argue about the merits of creativity here, or that someone putting a prompt into ChatGPT considers themselves an artist.
That's irrelevant. The work is created on OpenAI servers, by the LLMs hosted there, and is then distributed to whoever wrote the prompt.
Models run locally are distributed by whoever trained them.
If you train a model on whatever data you legally have access to, and produce something for yourself, it's one thing.
Distribution is where things start to get different.
You can sue the company making the LLM, which is what many have done.
The act of training by itself has been ruled to be fair use over and over again, including for LLMs, and there isn't much debate left there.
The test for infringement is if the output is transformative enough, and that is what NYT vs OpenAI etc. are arguing.
I agree there has to be a court case about it. I think the current argument, however, is that it is transformative, and therefore falls under fair use.
You don’t get to simply claim fair use based on how transformative your derivative work is.
“”” Section 107 calls for consideration of the following four factors in evaluating a question of fair use:
Purpose and character of the use, including whether the use is of a commercial nature or is for nonprofit educational purposes: Courts look at how the party claiming fair use is using the copyrighted work, and are more likely to find that nonprofit educational and noncommercial uses are fair. This does not mean, however, that all nonprofit education and noncommercial uses are fair and all commercial uses are not fair; instead, courts will balance the purpose and character of the use against the other factors below. Additionally, “transformative” uses are more likely to be considered fair. Transformative uses are those that add something new, with a further purpose or different character, and do not substitute for the original use of the work.
Nature of the copyrighted work: This factor analyzes the degree to which the work that was used relates to copyright’s purpose of encouraging creative expression. Thus, using a more creative or imaginative work (such as a novel, movie, or song) is less likely to support a claim of a fair use than using a factual work (such as a technical article or news item). In addition, use of an unpublished work is less likely to be considered fair.
Amount and substantiality of the portion used in relation to the copyrighted work as a whole: Under this factor, courts look at both the quantity and quality of the copyrighted material that was used. If the use includes a large portion of the copyrighted work, fair use is less likely to be found; if the use employs only a small amount of copyrighted material, fair use is more likely. That said, some courts have found use of an entire work to be fair under certain circumstances. And in other contexts, using even a small amount of a copyrighted work was determined not to be fair because the selection was an important part—or the “heart”—of the work.
Effect of the use upon the potential market for or value of the copyrighted work: Here, courts review whether, and to what extent, the unlicensed use harms the existing or future market for the copyright owner’s original work. In assessing this factor, courts consider whether the use is hurting the current market for the original work (for example, by displacing sales of the original) and/or whether the use could cause substantial harm if it were to become widespread. “””
https://www.copyright.gov/fair-use/
Yea, a finding that training is transformative would be pretty significant and it's likely that the precedent of thumbnail creation being deemed transformative would likely steer us towards such a finding. Transformative is always a hard thing to bank on because it is such a nebulous and judgement based call. There are excellent examples of how precise and gritty this can get in audio sampling.
Didn't know about thumbnails being fair use. In that case, I just don't see an argument that genAI training on source code is less transformative than thumbnails.
From the article:
> He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch.
From GPL2:
> The source code for a work means the preferred form of the work for making modifications to it. For an executable work, complete source code means all the source code for all modules it contains, plus any associated interface definition files, plus the scripts used to control compilation and installation of the executable.
Is a project's test suite not considered part of its source code? When I make modifications to a project, its test cases are very much a part of that process.
If the test suite is part of this library's source code, and Claude was fed the test suite or interface definition files, is the output not considered a work based on the library under the terms of LGPL 2.1?
It's transformative, so no.
Legally, using the tests to help create the reimplementation is fine.
However, it seems possible you can't redistribute the same tests under the MIT license. So the reimplementation MIT distribution could need to be source code only, not source code plus tests. Or, the tests can be distributed in parallel but still under LGPL, not MIT. It doesn't really matter since compiled software won't be including the tests anyways.
> It's transformative, so no.
I'm not following your logic there, and I don't see any mention of "transformative" in the license. Can you explain what you mean?
Sorry, I misspoke. Transformation is what makes the LLM itself legal -- its training data is sufficiently transformed into weights.
And so, a work being sufficiently transformative is one way in which copyright no longer applies, but that's not the case here specifically. The specific case here is essentially just a clean-room reimplementation (though technically less "clean", but still presumably the same legally). But the end result is still a completely different expression of underlying non-copyrightable ideas.
And in both cases, it doesn't matter what the original license was. If a resulting work is sufficiently transformative or a reimplementation, copyright no longer applies, so the license no longer applies.
That's interesting, but it misses my point:
The library's test suite and interfaces were apparently used directly, not transformed. If either of those are considered part of the library's source code, as the license's wording seems to suggest, then I think output from their use could be considered a work based on the library as defined in the license.
Legally that's been established as acceptable.
Google LLC v Oracle America assumed (though didn't establish) that API's are copyrightable... BUT that developing against them falls under fair use, as long as the function implementations are independent.
Test suites are again generally considered copyrightable... but the behavior being tested is not.
So no, it's not considered to be a work based on the library. This seems pretty clear-cut in US law by now.
Also, the LGPL text doesn't say "work based on the library". It says "If you modify a copy of the Library", and this is not a "combined work" either. And the whole point is that this is not a modified copy -- it's a reimplementation.
In theory, a license could be written to prevent running its tests from being run against software not derived from the original, i.e. clean-room reimplementations. In practice, it remains dubious whether any court would uphold that. And it would also be trivial to then get around it, by taking advantage of fair use to re-implement the tests in e.g. plain English (or any specification language), and then re-implementing those back into new test code. Because again, test behaviors are not copyrightable.
> Also, the LGPL text doesn't say "work based on the library".
It does, about a dozen times.
Are you perhaps referring to LGPL3? I think the relevant license under discussion here is LGPL2.1.
https://github.com/chardet/chardet/blob/6.0.0/LICENSE
I'm not well versed in copyright case law, so I won't argue with the rest of what you wrote. Thanks for elaborating on your thoughts.
Oh sorry, I was indeed looking at the wrong LGPL version. I stand corrected, thanks! But yes, all the same points stand. Good discussion!
But the tests were transformed to the new language, they are not copied as-is.
Software patents would work as you describe, but not copyright.
Google v. Oracle ruled that use of APIs are fair game and could be argued that test cases are strictly a use of APIs and not implementation.
Google vs Oracle ruled that APIs fall under copyright (the contrary was thought before). However, it was ruled that, in that specific case, fair use applied, because of interoperability concerns. That's the important part of this case: fair use is never automatic, it is assessed case by case.
Regarding chardet, I'm not sure "I wanted to circumvent the license" is a good way to argue fair use.
> Blanchard's account is that he never looked at the existing source code directly. He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch
This feels sort of like saying "I just blindly threw paint at that canvas on the wall and it came out in the shape of Mickey Mouse, and so it can't be copyright infringement because it was created without the use of my knowledge of Micky Mouse"
Blanchard is, of course, familiar with the source code, he's been its maintainer for years. The premise is that he prompted Claude to reimplement it, without using his own knowledge of it to direct or steer.
> Blanchard is, of course, familiar with the source code, he's been its maintainer for years.
I would argue it's irrelevant if they looked or didn't look at the code. As well as weather he was or wasn't familiar with it.
What matters is, that they feed to original code into a tool which they setup to make a copy of it. How that tool works doesn't really matter. Neither does it make a difference if you obfuscate that it's an copy.
If I blindfold myself when making copies of books with a book scanner + printer I'm still engaging in copyright infringement.
If AI is a tool, that should hold.
If it isn't "just" a tool, then it did engage in copyright infringement (as it created the new output side by side with the original) in the same way an employee might do so on command of their boss. Which still makes the boss/company liable for copyright infringement and in general just because you weren't the one who created an infringing product doesn't mean you aren't more or less as liable of distributing it, as if you had done so.
>that they feed to original code into a tool which they setup to make a copy of it
Well, no. They fed the spec (test cases, etc) into a tool which made a new program matching the spec. This is not a copy of the original code.
But also this feels like arguing over the color of the iceberg while the titanic sinks. If you have a tool that can make code to spec, what is the value in source code anymore? Even if your app is closed-source, you can just tell claude to write new code that does the same thing.
Blanchard fed the spec to the tool, and Anthropic fed the code to the tool, so Blanchard didn't do anything wrong, and Anthropic didn't do anything wrong. Nothing to see here.
> Blanchard fed the spec to the tool,
Yes...
> and Anthropic fed the code to the tool,
Presumably, as part of the massive amount of open-source code that must have been fed in to train their model.
> so Blanchard didn't do anything wrong, and Anthropic didn't do anything wrong. Nothing to see here.
This is meant as irony, right?
Everyone writes as if he just fed the spec and tests to Claude Code. Ignoring for now that the tests are under LGPL as well, the commit history shows that this has been done with two weeks of steering Claude Code towards the desired output. At every one of these interactions, the maintainer used his deep knowledge of the chardet codebase to steer Claude.
if the actual text of the code isn't the same or obviously derivative, copyright doesn't apply at all.
What does derivative mean here? Because IMO it means that the existing work was used as input. So if you used a LLM and it was trained on the existing work, that's a derivative work. If you rot13 encode something as input, so you can't personally read it, and then a device decides to rot13 on it again and output it, that's a derivative work.
In order for it to be creatively derivative you would need to copy the structure, logic, organization, and sequence of operations not just reimplement the functionality. It is pretty clear in this case that wasn't done.
It's not clear at all.
As a cynical person I assume all the frontier LLMs were trained on datasets that include every open source project, but as a thought experiment, if an LLM was trained on a dataset that included every open source project _execept_ chardet, do you think said LLM would still be able to easily implement something very similar?
There is no doubt in my mind that it could still do it.
Of course, the problem with this interpretation is that all modern LLMs are derivatives from huge amounts of text under completely different licenses, including "All rights reserved", and therefore can not be used for any purpose.
I'm not sure how you square the circle of "it's alright to use the LLM to write code, unless the code is a rewrite of an open source project to change its license".
> Of course, the problem with this interpretation is that all modern LLMs are derivatives from huge amounts of text under completely different licenses, including "All rights reserved", and therefore can not be used for any purpose.
> I'm not sure how you square the circle of "it's alright to use the LLM to write code
You seem like you're on the cusp of stating the obvious correct conclusion: it isn't.
> Because IMO it means that the existing work was used as input
That's your opinion (since you said "IMO"), not the actual legal definition.
LLMs do not encode nor encrypt their training data. The fact they can recite training data is a defect not a default. You can understand this more simply by calculating the model size as an inverse of a fantasy compression algorithm that is 50% better than SOTA. You'll find you'd still be missing 80-90% of the training data even if it were as much of a stochastic parrot as you may be implying. The outputs of AI are not derivative just because they saw training data including the original library.
Then onto prompting: 'He fed only the API and (his) test suite to Claude'
This is Google v Oracle all over again - are APIs copyrightable?
> LLMs do not encode nor encrypt their training data. The fact they can recite training data is a defect not a default.
About this specific point, it is unclear how much of a defect memorization actually is - there are also reasons to see it as necessary for effective learning. This link explains it well:
https://infinitefaculty.substack.com/p/memorization-vs-gener...
> This is Google v Oracle all over again - are APIs copyrightable?
Yes this is the best way to ask the question. If I take a public facing API and reimplement everything, whether it's by human or machine, it should be sufficient. After all, that's what Google did, and it's not like their engineers never read a single line of the Java source code. Even in "clean room" implementations, a human might still have remembered or recalled a previous implementation of some function they had encountered before.
I find the "compression" argument not very strong, both because copyright still applies to (very) lossy codecs (e.g. your 16kbps Opus file of Thriller infringes, even if the original 192khz/32bit wav file was 12,000kbps), and because copyright still applies to transformed derivative works (a tiny midi file of Thriller might still be enough for the Jackson's label to get you)
See also: https://monolith.sourceforge.net/, which seeks to ask the question:
> But how far away from direct and explicit representations do we have to go before copyright no longer applies?
Copyright protects even very abstract aspects of human creative expression, not just the specific form in which it is originally expressed. If you translate a book into another language, or turn it into a silent movie, none of the actual text may survive, but the story itself remains covered by the original copyright.
So when you clone the behavior of a program like chardet without referencing the original source code except by executing it to make sure your clone produces exactly the same output, you may still be infringing its copyright if that output reflects creative choices made in the design of chardet that aren't fully determined by the functional purpose of the program.
If you pirate a movie and reencode it, does that apply as well? You can still watch the movie and it is “obviously” the same movie, even though the bytes are completely different. Here you can use the program and it is, to the user, also the same.
> If it isn't "just" a tool, then it did engage in copyright infringement
Copyright infringement is a thing humans do. It's not a human.
Just like how the photos taken by a monkey with a camera have no copyright. Human law binds humans.
Correct. The human who shares the copy is the one who engages in copyright infringement.
So, let's say that rather than actually touching any copyrighted material, a human merely tells an AI about how to go onto the internet and find copyrighted material, download it, and ingest it for training. The AI, fully autonomously, does so, and after training itself on the material deletes it so no human ever downloads, consumes, or shares it.
If we are saying AI is "more than a tool", which seems to be the case courts are leaning since they've ruled AI output without direct human involvement is not copyrightable[0], then the above seems like it would be entirely legal.
[0] https://www.copyright.gov/newsnet/2025/1060.html
Someone would likely get prosecuted if they instructed AI agent to run say a pump and dump scheme...
Even if the final output doesn't have copyright protection it might still be copyright violation. I think it could be reasonable to have work that itself violates copyright when distributed even if it does not have copy right itself.
I just don't see how it's relevant whether he did look or didn't. In my opinion, it's not just legally valid to make a re-implementation of something if you've seen the code as long as it doesn't copy expressive elements. I think it's also ethically fine as well to use source code as a reference for re-implementing something as long as it doesn't turn into an exact translation.
It's actually not legally fine, or at least it's extremely dangerous. Projects that re-implement APIs presented by extremely litigious companies specifically do not allow people who, for instance, have seen the proprietary source code to then work on the project.
I don't think fear or legal action makes it illegal.
If I know it is legal to make a turn at a red light. And I know a court will uphold that I was in the right but a police officer will fine me regardless and I would need to go to actually pursue some legal remedy I'm unlikely to do it regardless of whether it is legal because it is expensive, if not in money but time.
In the case of copyright lawsuits they are notoriously expensive and long so even if a court would eventually deem it fine, why take the chance.
That's my point. It's dangerous and there are sharks in the water. That sounds like you're not going to have a good time if you do the described approach to someone who might assert you're infringing.
My understanding is that that is a maximalist position for the avoidance of risk, and is sufficient but probably not necessary.
Right. The alternative is that we reward Dan for his 14 years of volunteer maintenance of a project... by banning him from working on anything similar under a different license for the rest of his life.
Ignoring the legal or ethical concerns. Let’s say we live in a world where the cost of copying code is so close to zero that it’s indistinguishable from a world without copyright.
Anything you put out can and will be used by whatever giant company wants to use it with no attribution whatsoever.
Doesn’t that massively reduce the incentive to release the source of anything ever?
No, because (most) people don't work on OSS for vanity, they do it to help other people, whether it's individuals or groups of individuals, ie corporations.
It's the same question as, if an AI can generate "art", or photographers can capture a scene better than any (realistic) painter, then will people still create art? Obviously yes, and we see it of course after Stable Diffusion was released three years ago, people are still creating.
I don’t know what a world without copyright does to corporate sponsored open source. It certainly reduces it because there are many corporate sponsored projects that monetize through dual licensing. My guess is in a world where you can’t even guarantee attribution, it’s much harder to convince your boss to let you open source a project in the first place.
So ignoring people who are being paid by corporations directly to work on open source, in my experience the vast majority of contributors expect to be able to monetize their work eventually in a way that requires attribution. And out of the small number who don’t expect a monetary return of any kind, a still smaller number don’t expect recognition.
If this weren’t the case you’d see a much larger amount of anonymous contributions. There are people who anonymously donate to charity. The vast majority want some kind of recognition.
Obviously we still see art, if you greatly reduce the monetary benefit to producing art, you’ll see a lot less of it. This is especially true of non trivial open source software that unlike static artwork requires continual maintenance.
Most commercial software that I've used has the model of a legal moat around a pretty crappy database schema.
The non IP protection has largely been in the effort involved in replicating an application's behavior and that effort is dropping precipitously.
You must not have used much commercial software outside of crappy business SaaS.
Truth
Yes, and it reduces the incentives to release binaries too. Such a world will be populated by almost entirely SaaS, which can still compete on freedom.
Oracle had it's day in court with Google over the Java APIs. Reimplementing APIs can be done without copyright infringement, but Oracle must have tried to find real infringement during discovery.
In this case, we could theoretically prove that the new chardet is a clean reimplementation. Blanchard can provide all of the prompts necessary to re-implement again, and for the cost of the tokens anyone can reproduce the results.
>This feels sort of like saying "I just blindly threw paint at that canvas on the wall and it came out in the shape of Mickey Mouse, and so it can't be copyright infringement because it was created without the use of my knowledge of Micky Mouse"
IANAL, but that analogy wouldn't work because Mickey Mouse is a trademark, so it doesn't matter how it is created.
Can anyone find the actual quote where Blanchard said this?
My understanding was that his claim was that Claude was not looking at the existing source code while writing it.
Conveniently ignoring the likelihood that Claude had been trained on the freely accessible source code.
Does he have access to Claude's training data? How can he claim Claude wasn't trained on the original code?
Isn't this a red herring? An API definition is fair use under Google v. Oracle, but the test suite is definitely copyrightable code!
If you only stick to the API and ignore the implementation, it is not Mickey Mouse any more but a rodent. If it was just a clone it wouldn't be 50x as fast. Nevertheless, APIs apparently can be copyrightable. I generally disagree with this; it's how PC compatibles took off, giving consumers better options.
Wait what, didn't oracle lose the case against Google? Have I been living in an alternate reality where API compatibility is fair use?
> This feels sort of like saying "I just blindly threw paint at that canvas on the wall and
> He fed only the API and the test suite to Claude and asked it
Difference being Claude looked; so not blind. The equivalent is more like I blindly took a photo of it and then used that to...
Technically did look.
The article is poorly written. Blanchard was a chardet maintainer for years. Of course he had looked at it's code!
What he claimed, and what was interesting, was that Claude didn't look at the code, only the API and the test suite. The new implementation is all Claude. And the implementation is different enough to be considered original, completely different structure, design, and hey, a 48x improvement in performance! It's just API-compatible with the original. Which as per the Google Vs oracle 2021 decision is to be considered fair use.
did he claim that Claude wasn't trained on the original? Or just that he didn't personally provide Claude with a copy?
I recon the latter, how would he know what was in Claude's training data?
> What he claimed, and what was interesting, was that Claude didn't look at the code
Who opened the PR? Who co-authored the commits? It's clearly on Github.
> Blanchard was a chardet maintainer for years. Of course he had looked at its code!
So there you have it. If he looked, he co-authored then there's that.
If I put my signature on Picasso painting, it doesn't make me co-author of said painting.
Blanchard is very clear that he didn't write a single line of code. He isn't an author, he isn't a co-author.
Signing GitHub commit doesn't change that.
> Blanchard is very clear that he didn't write a single line of code
He used Claude to write it. Difference? The fact that I write on the notepad vs printed it out = I didn't do it?
> Signing GitHub commit doesn't change that.
That's the equivalent of me saying I didn't kill anyone. The fingerprints on the knife doesn't change that.
I'll take a commit authored by someone else and then git amend the author to myself, did I write that commit then? By your logic I did apparently.
What if we said that generative AI output is simply not copyrightable. Anything an AI spits out would automatically be public domain, except in cases where the output directly infringes the rights of an existing work.
This would make it so relicensing with AI rewrites is essentially impossible unless your goal is to transition the work to be truly public domain.
I think this also helps somewhat with the ethical quandary of these models being trained on public data while contributing nothing of value back to the public, and disincentivize the production of slop for profit.
We did in fact say so.
https://www.carltonfields.com/insights/publications/2025/no-...
> No Copyright Protection for AI-Assisted Creations: Thaler v. Perlmutter
> A recent key judicial development on this topic occurred when the U.S. Supreme Court declined to review the case of Thaler v. Perlmutter on March 2, 2026, effectively upholding lower court rulings that AI-generated works lacking human authorship are not eligible for copyright protection under U.S. law
> > A recent key judicial development on this topic occurred when the U.S. Supreme Court declined to review the case of Thaler v. Perlmutter on March 2, 2026, effectively upholding lower court rulings that AI-generated works lacking human authorship are not eligible for copyright protection under U.S. law
This was AI summary? Those words were not in the article.
The courts said Thaler could not have copyright because he refused to list himself as an author.
> This would make it so relicensing with AI rewrites is essentially impossible unless your goal is to transition the work to be truly public domain.
That's not true at all. Anyone could follow these steps:
1. Have the LLM rewrite GPL code.
2. Do not publish that public domain code. You have no obligation to.
3. Make a few tweaks to that code.
4. Publish a compiled binary/use your code to host a service under a proprietary license of your choice.
In the corporate world, we've started using reimplementation as a way to access tooling that security won't authorize.
Sec has a deny by default policy. Eng has a use-more-AI policy. Any code written in-house is accepted by default. You can see where this is going.
We've been using AI to reimplement tooling that security won't approve. The incentives conspired in the worst outcome, yet here we are. If you want a different outcome, you need to create different incentives.
Not Invented Here's long, slow mutagenic march toward full antibiotic resistance continues apace.
There is a fundamental corpo-cognitive dissonance, to boot. If "AI" is cheap enough and good enough to implement security-relevant software from `git init` repeatedly, why isn't it also cheap enough and good enough to assess and approve the security of third-party software at pace with internal adoption? Is there some basis to believe LLMs' leverage on production differs from its leverage on analysis of existing code?
I largely agree with the author that AI can't just magically remove license agreements by rewriting code.
However, I take issue with his version of history:
>The history of the GPL is the history of licensing tools evolving in response to new forms of exploitation: GPLv2 to GPLv3, then AGPL.
GPLv3 set open source backwards: it wasn't an evolution to protect anything, it was a an overly paranoid failure. Don't believe me? Just count how many GPL3 vs. how many GPL2 projects have been started since GPL3 dropped.
Again, I'm very pro-OSS, but let's not pretend the community has always had a straight line of progress forward; some stuff is crazy Stallman stuff that set us back.
If Blanchard is claiming not to have been substantively involved in the creation of the new implementation of chardet (i.e. "Claude did it"), then the new implementation is machine generated, and in the USA cannot be copyright and thus cannot be licensed.
If he is claiming to have been somehow substantively "enough" involved to make the code copyrightable, then his own familiarity with the previous LGPL implementation makes the new one almost certainly a derivative of the original.
>then his own familiarity with the previous LGPL implementation makes the new one almost certainly a derivative of the original.
The "clean room rewrite" is just an extreme way to have a bulletproof shield against litigation. Not doing it that way doesn't automatically make all new code he writes derivative solely because he saw how the code worked previously.
If the clean room re-write was done entirely by Claude, then the result cannot be copyright in the USA, and thus there is no license at all.
And if he was in fact more involved (which he appears to deny) that it's a bit weak to say that someone with huge familiarity with chardet could choose to reimplement chardet without the result being derivative.
So if I read any LGPL code in my life, I can never think about working on something similar in my life?
There's a difference between "I've read a LGPL code once, maybe I could do something similar" and "I've been reading this LGPL code for 12 years and now I'm going to do exactly the same thing".
I agree with the thrust of this article, that norms and what we perceive as good or desirable extend considerably beyond the minimum established by law.
But a point that was not made strongly, which highlights this even more, is that this goes in every direction.
If this kind of reimplementation is legal, then I can take any permissive OSS and rebuild it as proprietary. I can take any proprietary software and rebuild it as permissive. I can take any proprietary software and rebuild it as my own proprietary software.
Either the law needs to catch up and prevent this kind of behavior, or we're going to enter an effectively post-copyright world with respect to software. Which ISN'T GOOD, because that will disincentivize any sort of open license at all, and companies will start protecting/obfuscating their APIs like trade secrets.
It goes in one direction only.
Companies can take open-source software and make a proprietary reimplementation. You can't take a proprietary software and make an open source GPL version.
I am absolutely certain that if you tried you would be sued to oblivion. But big company screwing up open source is not even news anymore. In fact I (still) believe that the fact that even though LLMs were trained on tons of GPL and AGPL or even unlicensed software it's considered ok to use LLM code in proprietary projects is example of just that.
From a strictly legal perspective the two are equivalent. The fact that there are structural injustices in the system is true, but that's not a question that any answer to "what should be legal" can fix.
I've been thinking this for over two years, that's why I stopped contributing to open source at that time - my work was only gonna be exploited to make rich people richer regardless of the license.
Crazy that only now we're seeing a bunch of articles coming to the same conclusion now.
I think copyright should still apply, but if it doesn't, we need new laws - ones which protect all human work, creative or not. Laws should serve and protect people, not algorithms and not corporations "owning" those algorithms.
I put owning in quotes because ownership should go to the people who did the work.
Buying/selling ownership of both companies and people's work should be illegal just like buying/selling whole humans is. Even if it took thousands of years to get here.
Money should not buy certain things because this is the root cause of inequality. Rich people are not getting richer at a faster rate by being more productive than everyone else but by "owning" other people's work and using it as leverage to extract even more from others.
Maybe LLM and mass unemployment of white collar workers will be the wakeup call needed for a reform. Or revolution.
Last time this happened was during the second industrial revolution and that's how communism got popular. We should do better this time because this is the last revolution which might be possible.
Wow, it feels like this argument rewired my brain.
When I first read about the chardet situation, I was conflicted but largely sided on the legal permissibility side of things. Uncomfortably I couldn't really fault the vibers; I guess I'm just liberal at heart.
The argument from the commons has really invoked my belief in the inherent morality of a public good. Something being "impermissible" sounds bad until you realize that otherwise the arrow of public knowledge suddenly points backwards.
Seeing this example play out in real life has had retroactive effects on my previously BSD-aligned brain. Even though the argument itself may have been presented before, I now understand the morals that a GPL license text underpins better.
FWIW I like to explain it to folks like this: ignore all of your moral baggage around licensing and just focus on the fact that licensing is a legal tool of art that pretty much only becomes relevant in the context of threatening lawsuits.
BSD-type stuff is very simple because it says "here is this stuff. you can use it as long as you promise not to sue me. I promise not to sue you too."
Very simple.
GPL-type stuff is intrinsically more complex because it's trying to use the threatening power of lawsuits, to reduce overall IP lawsuits. So it has to say "Here is this stuff. You can use it as long as you promise not to sue me. I am only going to sue you, if you start pretending like you have the right to sue other folks over this stuff or anything you derive from it. You don't have the right to sue others for it, I made it, so please stop pretending and let's stop suing each other over this sort of thing."
Getting the entire legal nuance around that sort of counterfactual "I will only sue you if you try to pretend that you can sue others" is why they're more complex. And the simplest copyleft licenses like the Mozilla Public License have a very rigid notion of what "the software" is, so like for MPL it's "this file is gonna never be used in a lawsuit, you can edit it ONLY as long as you agree that this file must never be used by you to sue someone else, if you try to mutate it in a way that lets you sue someone else then that's against our agreement and we reserve the right to sue you."
Whereas for GPL it's actually kind of nebulous what "the software" is -- everything that feeds into the eventual compiled binary, basically -- and so the license itself needs to be a little bit airy-fairy, "let's first talk about what conveying the software means...", in various ways.
The interesting thing here is that as far as the courts are initially ruling, these from-scratch reimplementations are not human works and therefore are not copyrightable, which makes them all kind of public domain. Slapping the MIT license on it was an overstep. If that's how things go then Free Software has actually won its greatest sweep with LLM ubiquity.
This is only worth arguing about because software has value. Putting this in context of a world where the cost of writing code is trending to 0, there are two obvious futures:
1. The cost continues to trend to 0, and _all_ software loses value and becomes immediately replaceable. In this world, proprietary, copyleft and permissive licenses do not matter, as I can simply have my AI reimplement whatever I want and not distribute it at all.
2. The coding cost reduction is all some temporary mirage, to be ended soon by drying VC money/rising inference costs, regulatory barriers, etc. In that world we should be reimplementing everything we can as copyleft while the inferencing is good.
There’s an other option. The cost of copying existing software trends to 0, but the cost of writing new software stays far enough above 0 that it is still relatively expensive.
There was a recent ruling that LLM output is inherently public domain (presumably unless it infringes some existing copyright). In which case it's not possible to use them to "reimplement everything we can as copyleft".
it's more complicated, the ruling was that AI can't be an author and the thing in question is (de-facto) public domain because it has no author in context of the "dev" claim it was fully build by AI
but AI assisted code has an author and claiming it's AI assisted even if it is fully AI build is trivial (if you don't make it public that you didn't do anything)
also some countries have laws which treat it like a tool in the sense that the one who used it is the author by default AFIK
The value of software has never been tied to the cost of writing it, even if you don't distribute it your still breaking the law.
The article is proceeding from the premise that a reimplementation is legal (but evil). To help my understanding of your comment, do you mean:
1. An LLM recreating a piece of software violates its copyright and is illegal, in which case LLM output can never be legally used because someone somewhere probably has a copyright on some portion of any software that an LLM could write.
2. You read my example as "copying a project without distributing it", vs. "having an LLM write the same functionality just for me"
There will always be cost though. Even if perfect code is getting one-shotted out, that is constantly maintained and adapted to changing conditions and technology, it simply can't stay at 0 forever because one day the power is surely going to go out!
More and more I am drawn to these kinds of ideas lately, perhaps as a kind of ethical sidestep, but still:
- https://wiki.xxiivv.com/site/permacomputing.html
- https://permacomputing.net/
It's not going to solve any general issue here, but the one thing these freaks need that can't be generated by their models is energy, tons of it. So, the one thing I can do as an individual and in my (digital) community is work to be, in a word, self-sustainable. And depending on my company I guess, if I was a CEO I would hope I was wise enough to be thinking on the same lines.
Everyone is making beautiful mountains from paper and wire. I will just be happy to make a small dollhouse of stone, I think it will be worth it. How can we see not just at least some small-level of hubris otherwise?
I'm probably spitting in the wind, but stuff like this is why I removed all my hosted open source projects. I manage several niche projects that I have now converted to binary only releases (to almost no push back). It's niche enough that it's not very hard to get LLMs to output chunks of code that it managed to scrape before I took it offline. I don't see many people talking about this angle, but LLMs ripping off my work killed my open source efforts.
This take, which I've seen in a few different places now, seems 100% bonkers. A world where anyone can cheaply reimplement anyone else's software and use it on hardware of their own choosing in their own designs and for their own purposes is a free software utopia.
This isn't a problem, this is the goal. GNU was born when RMS couldn't use a printer the way he wanted because of an unmodifiable proprietary driver. That kind of thing just won't happen in the vibe coded future.
Surprised they don't mention Google LLC v. Oracle America, Inc. Seems a bit myopic to condone the general legality while arguing "you can only use it how I like it".
It also doesn't talk about the far more interesting philosophical queston. Does what Blanchard did cover ALL implementations from Claude? What if anyone did exactly what he did, feed it the test cases and say "re-implement from scratch", ostensibly one would expect the results to be largely similar (technically under the right conditions deterministically similar)
could you then fork the project under your own name and a commercial license? when you use an LLM like this, to basically do what anyone else could ask it to do how do you attach any license to it? Is it first come first serve?
If an agent is acting mostly on its own it feels like if you found a copy of Harry Potter in the fictional library of Babel, you didn't write it, just found it amongst the infinite library, but if you found it first could you block everyone else that stumbles on a near-identical copy elsewhere in the library? or does each found copy represent a "Re-implementation" that could be individually copyrighted?
It should be noted that the Rust community is also guilty of something similar. That is, porting old GPL programs, typically written in C, to Rust and relicensing them as MIT.
> porting old GPL programs, typically written in C, to Rust and relicensing them as MIT
Everything for memory safety.
If an AI can license-wash open source software like this then the licenses become meaningless. Which is fascinating. Commercial software cloning that is simple enough for an average person to drive is next and the ultimate form of piracy, see an app for $10? Don’t fancy paying? Just ask ChatGPT for a clone. Future is going to be wild.
You've just described why every SaaS stock has taken a beating in the last 6 months.
How long until this thing is good enough to clone photoshop? Or Skyrim? I think all bets are off for the software world.
> ultimate form of piracy
Nothing was stolen, not even copied, lamest piracy I've heard of.
I take your point, but if the re-implementation looks the same, I would say it’s a form of copying. (Which I don’t think is a problem, I don’t think you should be able to own sequences of numbers.)
I think the direction we are going, the GPL is going to fade away. I think people will look at this like writing a book and claiming the ideas in the book cannot be copied. This debate is not that different from the ones going on in the music industry. I open sourced my latest software as Apache 2.0 after debating a lot about this. Unless the FSF wins in court in the next <=2-3 years, there is no coming back from this.
There's a Japanese version of that page, written in classical text writing direction, in columns. Which is cool. Makes me wonder, though - how readable is it with so many English loanwords which should be rotated sideways to fit into columns?
Total digression but yeah, that layout is stupid and the way those words are dropped in using Romaji makes no sense. That's not how Japanese people lay out pages on the web. In fact I don't think I've ever seen a Japanese web page laid out like a book like this, and in general I'd expect the English proper nouns and words that don't have obvious translations to get transliterated into Katakana. Smells like automatic conversion added by someone not really familiar with common practices for presenting Japanese on the web.
He also has a Korean vertical layout that lays out Latin-character words the same way. Is this common in Korea when vertical layout is used? The author seems to be Korean.
Looks like Wikipedia has an example of Traditional Chinese vertical layout with the Latin letters rotated as in TFA's layout (https://en.wikipedia.org/wiki/Horizontal_and_vertical_writin...)
It doesn't matter if it's legitimate. The people that use it don't care. They just find it online and click download. This is the reality.
It seems that this chap didn't go and implement a new library, he reimplemented an existing one and became sole-controller of it. i.e. he seems to have taken its reputation, brand whatever you call it away from the contributors and entirely to himself. Their work of establishing it as a well known solution is no longer recognised.
So of course we feel that something wrong has happened even if it's not easy to put one's finger on it.
"Antirez closes his careful legal analysis as though it settles the matter. Ronacher acknowledges that “there is an obvious moral question here, but that isn't necessarily what I'm interested in.” Both pieces treat legal permissibility as a proxy for social legitimacy. "
This whole article is just complaining that other people didn't have the discussion he wanted.
Ronacher even acknowledged that it's a different discussion, and not one they were trying to have at the moment.
If you want to have it, have it. Don't blast others for not having it for you.
Having this discussion involves blasting others for not considering it. Consider the rest of the paragraph you quoted:
> But law only says what conduct it will not prevent—it does not certify that conduct as right. Aggressive tax minimization that never crosses into illegality may still be widely regarded as antisocial. A pharmaceutical company that legally acquires a patent on a long-generic drug and raises the price a hundredfold has not done something legal and therefore fine. Legality is a necessary condition; it is not a sufficient one.
If the discussion inherently cannot be had without blasting innocent bystanders, I don't think it's a discussion worth having.
It might even be morally abhorrent to have such a discussion in the first place!
Broadly speaking, the “freedom of users” is often protected by competition from competing alternatives. The GNU command line tools were replacements for system utilities. Linux was was a replacement for other Unix kernels. People chose to install them instead of proprietary alternatives. Was it due to ideology or lower cost or more features? All of the above. Different users have different motivations.
Copyleft could be seen as an attempt to give Free Software an edge in this competition for users, to counter the increased resources that proprietary systems can often draw on. I think success has been mixed. Sure, Linux won on the server. Open source won for libraries downloaded by language-specific package managers. But there’s a long tail of GPL apps that are not really all that appealing, compared to all the proprietary apps available from app stores.
But if reimplementing software is easy, there’s just going to be a lot more competition from both proprietary and open source software. Software that you can download for free that has better features and is more user-friendly is going to have an advantage.
With coding agents, it’s likely that you’ll be able to modify apps to your own needs more easily, too. Perhaps plugin systems and an AI that can write plugins for you will become the norm?
> Was it due to ideology or lower cost or more features?
It was due to access.
You can't put a copyright and MIT license on something you generated with AI. It is derived from the work of many unknown, uncredited authors.
Think about it; the license says that copies of the work must be reproduced with the copyright notice and licensing clauses intact. Why would anyone obey that, knowing it came from AI?
Countless instances of such licenses were ignored in the training data.
When learning is sufficiently atomized and recombined, creations cease to be "derived from" in a legal sense.
A lego sculpture is copyrighted. Lego blocks are not. The threshold between blocks and sculpture is not well-defined, but if an AI isn't prompted specifically to attempt to mimic an existing work, its output will be safely on the non-copyrighted side of things.
A derivative work is separately copyrightable, but redistribution needs permission from the original author too. Since that usually won't be granted or would be uneconomical, the derivative work can't usually be redistributed.
AI-produced material is inherently not copyrightable, but not because it's a derivative work.
Token prediction is a form of "learning" that is reinforced by the goal of reproducing the correct next token of the work, rather that acquiring ideas and concepts. For instance, given the prefix "Four score and seven years", the weights are adjusted until "ago" is correctly predicted, which is a fancy way of saying that it was stored in the model in a lossy way. The model "learned" that "ago" follows "four score and seven years" exactly the way your hard drive "learns" the audio and video frames of a movie when you download a .mp4 file.
I dispute the idea that token sequences reproduced from the model are not derived works.
I predict, no pun intended, that a time is coming when the idea that it's not a derived work will be challenged in mainstream law.
The slop merchants are getting a free ride for the time being.
Only because that quote is famous.
As you said, it's lossy. Try it with any other distinctive but non-famous passage, and you won't get a correct prediction for the immediately following clause, much less for multiple sentences or paragraphs.
That's the case even when an LLM correctly identifies which book the prompted text is from. It still won't accurately continue on from some arbitrary passage. By the time you ask it to reproduce hundreds of words, you're into brand new book territory. Even when it's slop content, it's distinct slop.
The exceptions are cases where a significant number of humans would also know a particular quote from memory. Then, chances are, a frontier LLM will too.
You know how else you can reproduce a quote? Search for it on google, and search the resulting top hits; if it's a significant quote, multiple people have probably quoted it -- legally. You can also search a pirate library for the actual book, and search the book for the quote; while illegal, it's very simple to do, so unless you propose to make the free and open internet illegal, I'd suggest that banning LLMs for being "derivative work" creation engines is not so different from destroying the internet.
> I predict, no pun intended, that a time is coming when the idea that it's not a derived work will be challenged in mainstream law.
If judges have any sense whatsoever, LLM generations (without specific prompt crafting to mimic existing works) will be judged to not be derived works and therefore not be violating copyright, in the same sense that you can live and breathe Taylor Swift's music, create new music in the same style, and still not be violating copyright.
The Stability AI case, and how Judge Orrick deals with it, will be interesting and uninteresting at the same time. It deals primarily with the fact that after specific prompting, an image-generation AI can generate something fairly close to existing copyrighted images. That doesn't say anything more about whether LLMs are inherently producers of [only or primarily] derivative works, just as the fact that a human can violate copyright doesn't say anything about whether humans primarily or exclusively output derivative works.
More likely, perhaps, is that everything will be so infused with LLM output that copyright ceases to be relevant, or forces copyright law to be rewritten from the ground up.
Copyright requirements, even prior to LLMs, weren't well-specified. There's no objective threshold for how close something has to be to a previous work before the new one violates copyright. It's whatever a judge thinks, refering to the 4-factor test but ultimately making subjective judgements about each of those prongs. It's all a house of cards, and LLMs may just be what topples it.
Courts have already ruled that AI-generated work belongs to the public domain. So, even the MIT license does not apply.
This article is setting up a bit of a moving target. Legal vs legitimate is at least only a single vague question to be defined but then the target changes to “socially legitimate” defined only indirectly by way of example, like aggressive tax avoidance as “antisocial”— and while I tend to agree with that characterization my agreement is predicated on a layering of other principals.
The fundamental problem is that once you take something outside the realm of law and rule of law in its many facets as the legitimizing principal, you have to go a whole lot further to be coherent and consistent.
You can’t just leave things floating in a few ambiguous things you don’t like and feel “off” to you in some way- not if you’re trying to bring some clarity to your own thoughts, much less others. You don’t have to land on a conclusion either. By all means chew over things, but once you try to settle, things fall apart if you haven’t done the harder work of replacing the framework of law with that of another conceptual structure.
You need to at least be asking “to what ends? What purpose is served by the rule?” Otherwise you’re stuck in things where half the time you end up arguing backwards in ways that put purpose serving rules, the maintenance of the rule with justifications ever further afield pulled in when the rule is questioned and edge cases reached. If you’re asking, essentially, “is the spirit of the rule still there?” You’ve got to stop and fill in what that spirit is or you or people that want to control you or have an agenda will sweep in with their own language and fill the void to their own ends.
> If source code can now be generated from a specification, the specification is where the essential intellectual content of a GPL project resides. Blanchard's own claim—that he worked only from the test suite and API without reading the source—is, paradoxically, an argument for protecting that test suite and API specification under copyleft terms.
This is an interesting reversal in itself. If you make the specification protected under copyright, then the whole practice of clean room implementations is invalid.
Well, the license change sounds pretty strange, but to be honest if I were to use this software I would use it without adhering to the MIT. It's machine-created content which is not, in general, copyrightable. You can assert whatever license you want on such content, but I am not going to adhere to it. For example, I declare you may use the following under the Elastic License
> When GNU reimplemented the UNIX userspace, the vector ran from proprietary to free. Stallman was using the limits of copyright law to turn proprietary software into free software. […] The vector in the chardet case runs the other way.
That’s just your subjective opinion which many other people would disagree. I bet Armin Ronacher would agree that an MIT licensed library is even freer than an LGPL licensed library. To them, the vector is running from free to freer.
> Blanchard's account is that he never looked at the existing source code directly.
That’s a weird statement while releasing the new version of the same project. Maybe just release it as a new project, chardet-ai v1.0 or whatever.
It also erodes copyright. A decent amount of commercial software can be AI cloned with no copyright violation.
A lot of SaaS too, especially if AI can run a simple deploy.
We might be approaching a huge deflationary catastrophe in the cost of a lot of software. It’s not a catastrophe for the consumer but it is for the industry.
It's clear that we're entering a new era of copyright _expectations_ (whether we get new _legislation_ is different), but for now realise this: the people like me who like copyleft can do this too. We can take software we like, point an agent at it, and tell it to make a new version with the AGPL3.0-or-later badge on the front.
But the LLM contributions would likely be ruled public domain, so AGPL may not be enforceable on these.
The point of GPL is to restrict distribution. If there’s already an MIT version, it’s useless.
> The point of GPL is to restrict distribution.
no, it isn't. The point of the GPL is to grant users of the software four basic freedoms (run, study, modify and redistribute). There's no restriction to distribution per se, other than disallowing the removal of these freedoms to other users.
but the point of an EULA is to restrict distribution, so AGPL3 can help there.
IMHO, the API and Test Suite, particularly the latter, define the contract of the functional definition of the software. It almost doesn't matter what that definition looks like so long as it conforms to the contract.
There was an issue where Google did something similar with the JVM, and ultimately it came down to whether or not Oracle owned the copyright to the header files containing the API. It went all the way to the US supreme court, and they ruled in Google's favour; finding that the API wasn't the implementation, and that the amount of shared code was so minimal as to be irrelevant.
They didn't anticipate that in less than half a decade we'd have technology that could _rapidly_ reimplement software given a strong functional definition and contract enforcing test suite.
without discussing copyright, I don't believe any of this is copied. Which I think should be the argument that actually matters.
I downloaded both 6.0 and 7.0 and based on only a light comparison of a few key files, nothing would suggest to me that 7.0 was copied from 6.0, especially for a 41x faster implementation. It is a lot more organized and readable in my armature opinion, and the code is about 1/10th the size.
Not a lawyer, but my understanding is: In theory, copyright only protects the creative expression of source code; this is the point of the "clean room" dance, that you're keeping only the functional behavior (not protected by copyright). Patents are, of course, an entirely different can of worms. So using an LLM to strip all of the "creative expression" out of source code but create the same functionality feels like it could be equivalent enough.
I like the article's point of legal vs. legitimate here, though; copyright is actually something of a strange animal to use to protect source code, it was just the most convenient pre-existing framework to shove it in.
> this is the point of the "clean room" dance
which is the actual relevant part: they didn't do that dance AFIK
AI is a tool, they set it up to make a non-verbatim copy of a program.
Then they feed it the original software (AFIK).
Which makes it a side by side copy, as in the original source was used as reference to create the new program. Which tend to be seen as derived work even if very different.
IMHO They would have to:
1. create a specification of the software _without looking at the source code_, i.e. by behavior observation (and an interface description). I.e. you give the AI access to running the program, but not to looking into the insides of it. I really don't think they did it as even with AI it's a huge pain as you normally can't just brute force all combinations of inputs and instead need to have a scientific model=>test=>refine loop (which AI can do, but can take long and get stuck, so you want it human assisted, and the human can't have inside knowledge about the program).
2. then generate a new program from specification, And only from it. No git history, no original source code access, no program access, no shared AI state or anything like that.
Also for the extra mile of legal risk avoidance do both human assisted and use unrelated 3rd parties without inside knowledge for both steps.
While this does majorly cut cost of a clean room approach, it still isn't cost free. And still is a legal mine field if done by a single person, especially if they have enough familiarity to potentially remember specific peaces of code verbatim.
Is your (1) description of clean room implementation, and (2) description of what was done, actually correct?
(1): my understanding was that a party _with access to copyrighted material_ made the functional spec, which was communicated to a party without access [1]. Under my understanding, theres no requirement for the authors of the functional spec to be 'clean'.
(2) Afaict, they limited the AI to access of just the functional spec and audited that it did not see the original source.
Edit: Not sure if sharing the 'test suite' matters, probably something for the courts in the unlikely event this ever gets there.
[1] Following the definition of clean room re implementation as it relates to US precedent, ie that described in the wikipedia page.
Well sure they didn't do the dance, but you don't have to do the dance. The reason to do it is that it's a good defense in a lawsuit. Like you say, all of this is a legal minefield.
So my understanding was that the original code was specifically not fed into Claude. But was almost certainly part of its training data, which complicates things, but if that's fair use then it's not relevant? If training's not fair use and taints the output, then new-chardet is a derivative of a lot of things, not just old-chardet...
This is all new legal ground. I'm not sure if anyone will go to court over chardet, though, but something that's an actual money-maker or an FSF flagship project like readline, on the other hand, well that's a lot more likely.
> Then they feed it the original software (AFIK).
My understanding is they did do the dance. From the article: "He fed only the API and the test suite to Claude and asked it to reimplement the library from scratch."
One could still make the argument that using the test suite was a critical contributing factor, but it is not a part of the resulting library. So in my uninformed opinion, it seems to me like the clean room argument does apply.
Why are people even having problems with sharing their changes to begin with? Just publishing it somewhere does not seem too expensive. The risk of accidentally including stuff that is not supposed to become public? Or are people regularly completely changing codebases and do not want to make the effort freely available, maybe especially to competitors? I would have assumed that the common case is adding a missing feature here, tweaking something there, if you turn the entire thing on its head, why not have your own alternative solution from scratch?
> an argument for protecting that test suite and API specification under copyleft terms.
If we protect API under copyright, it makes it easier to prevent interoperability. We obviously do NOT want that. It would give big companies even more power.
Now in the US, the Supreme Court that the output of an LLM is not copyrightable. So even a permissive licence doesn't work for that reimplementation: it should be public domain.
Disclaimer: I am all for copyleft for the code I write, but already without LLMs, one could rewrite a similar project and use the licence they please. LLMs make them faster at that, it's just a fact.
Now I wonder: say I vibe-code a library (so it's public domain in the US), I don't publish that code but I sell it to a customer. Can I prevent them from reselling it? I guess not, since it's public domain?
And as an employee writing code for a company. If I produce public domain code because it is written by an LLM, can I publish it, or can the company prevent me from doing it?
Why does anyone need his new library? They can do what he did and make their own.
I'm glad we can fork things at a point and thumb our noses at those who wish to cash in on other's work.
Why would I make my own? The new library is released under MIT license and faster than the old one.
If you decide to improve it in any way to fit your needs you can merely tell your own AI to re-implement it with your changes. Then it's proprietary to you.
i've been following this for a while.. and the trend for copyright (of any form - books code pictures music whatever) being laundered by reinventing the "same" thing in-some-way.. is kind-of clear.
But what happens with the new things? Has the era of software-making (or creating things at large) finished, and from now on everything will be re-(gurgitated|implemented|polished) old stuff?
Or all goes back to proprietary everything.. Babylon-tower style, noone talks to noone?
edit: another view - is open-source from now on only for resume-building? "see-what-i've-built" style
Both sides are wrong on this actually. Computer generated code has no copyright protection.
>The U.S. Copyright Office (USCO) and federal courts have consistently ruled that AI-generated works—where the expressive elements are determined by the machine, even in response to a human prompt—lack the necessary human creative input and therefore cannot be copyrighted.
All this code is public domain. Your employees can publish "your" AI generated code freely and it won't matter how many tokens you spent generating it. It is not covered by copyright.
I feel like the licenses that suffer the most isn't the GPL, but the ones like SSPL. If your code can be re-implemented easily and legally by AWS using an LLM, why risk publishing it?
It does feel like open source is about to change. My hunch is that commercial open source (beyond the consultation model) risks disappearing. Though I'd be happy to be proven wrong.
Buried in here: Mark Pilgrim suddenly reappearing after his sudden disappearance years ago! Has he been up to anything since then?
I don't think this part is correct: "If you distribute modified code, or offer it as a networked service, you must make the source available under the same terms."
That's what something like AGPL does.
One of the things that irks me about this whole thing is, if it’s so clean room and distinct, why make the changes to the existing project? Why not make an entirely new library?
The answer to that, I think, is that the authors wanted to squat an existing successful project and gain a platform from it. Hence we have news cycle discussing it.
Nobody cares about a new library using AI, but squash an existing one with this stuff, and you get attention. It’s the reputation, the GitHub stars, whatever
I mean, Blanchard was the longtime maintainer of chardet already, and had wanted to relicense it for years. So I think that complicates your picture of "squatting an existing successful project".
Honestly it's a weird test case for this sort of thing. I don't think you'd see an equivalent in most open source projects.
I agree. But you can't copyright goodwill and reputation. Trademark does provide some protection there, right?
A lot of untagged IANAL takes here today.
I'm less concerned about AI eroding copyleft and more exited about AI eroding copy right.
See also "A Declaration of the Independence of Cyberspace" (https://www.eff.org/cyberspace-independence), and what a goofy, naive, misguided disaster that early internet optimism turned into.
No, AI does not mean the end of either copyright or copyleft, it means that the laws need to catch up. And they should, and they will.
Someone be brave, and do this to ZFS. Poke the Oracle bear!
1) Legality and morality are obviously different and unrelated concepts. More people should understand that.
2) Copyright was the wrong mechanism to use for code from the start, LLMs just exposed the issue. The thing to protect shouldn't be creativity, it should be human work - any kind of work.
The hard part of programming isn't creativity, it's making correct decisions. It's getting the information you need to make them. Figuring out and understanding the problem you're trying to solve, whether it's a complex mathematical problem or a customer's need. And then evaluating solutions until you find the right one. (One constrains being how much time you can spend on it.)
All that work is incredibly valuable but once the solution exists, it's each easier to copy without replicating or even understanding the thought process which led to it. But that thought process took time and effort.
The person who did the work deserved credit and compensation.
And he deserves it transitively, if his work is used to build other works - proportional to his contribution. The hard part is quantifying it, of course. But a lot of people these days benefit from throwing their hands up and saying we can't quantify it exactly so let's make it finders keepers. That's exploitation.
3) Both LLM training and inference are derivative works by any reasonable meaning of those words. If LLMs are not derivative works of the training data then why is so much training data needed? Why don't they just build AI from scratch? Because they can't. They just claim they found a legal loophole to exploit other people's work without consent.
I am still hoping the legal people take time to understand how LLMs work, how other algorithms, such as synonym replacement or c2rust work, decide that calling it "AI" doesn't magically remove copyright and the huge AI companies will be forced to destroy their existing models and train new ones which respect the licenses.
Imagine if the author has his way, and when we have AI write software, it becomes legally under the license of some other sufficiently similar piece of software. Which may or may not be proprietary. "I see you have generated a todo app very similar to Todoist. So they now own it." That does not seem like a good path either for open source software or for opening up the benefits of AI generated software.
Perhaps software patents may play an even bigger role in the future.
Or, hopefully, even less of a role.
LPGL is dead, long live the AI rewrites of your barely open source code
Easy solution for now:
Add something like this to NEW gpl /bsd/mit licenses:
'you are forbidden from reimplementing it with AI'
or just:
'all clones, reimpletetions with ai etc must still be GPL'
What if someone doesn't declare that it has been reimplemented using an LLM? Isn't it enough to simply declare that you have reimplemented the software without using an LLM? Good luck proving that in court...
One thing is certain, however: copyleft licenses will disappear: If I can't control the redistribution of my code (through a GPL or similar license), I choose to develop it in closed source.
Arguably, the GPL has always been the wrong choice if you want to authoritatively control redistribution.
That's a non-sequitur. chardet v7 is GPL-derived work (currently in clear violation of the GPL). If xe wanted it to be a different thing xe should've published as such. Simple as.
If the model wasn't trained on copyleft, if he didn't use a copyleft test suite and if he wasn't the maintainer for years. Clearly the intent here is copyright infringement.
If you have software your testsuite should be your testsuite, you do dev with a testsuite and then mit without releasing one. Depending on the test-suite it may break clean room rules, especially for ttd codebases.
I think what is happening is the collapse of the “greater good”. Open source is dependent upon providing information for the greater good and general benefit of its readers. However now that no one is reading anything, its purpose is for the great good of the most clever or most convincing or richest harvester.
shall we now have to think about the tradeoffs in adopting
- proprietary
- free
- slop-licensed
software?
We should just use LLMs to free more software and HW. Make it work against the system.
> Ronacher notes this as an irony and moves on. But the irony cuts deeper than he lets on. Next.js is MIT licensed. Cloudflare's vinext did not violate any license—it did exactly what Ronacher calls a contribution to the culture of openness, applied to a permissively licensed codebase. Vercel's reaction had nothing to do with license infringement; it was purely competitive and territorial. The implicit position is: reimplementing GPL software as MIT is a victory for sharing, but having our own MIT software reimplemented by a competitor is cause for outrage. This is what the claim that permissive licensing is “more share-friendly” than copyleft looks like in practice. The spirit of sharing, it turns out, runs in one direction only: outward from oneself.
This argument makes no sense. Are they arguing that because Vercel, specifically, had this attitude, this is an attitude necessitated by AI, reimplementation, and those who are in favor of it towards more permissive licenses? That certainly doesn't seem to be an accurate way to summarize what antirez or Ronacher believe. In fact, under the legal and ethical frameworks (respectively) that those two put forward, Vercel has no right to claim that position and no way to enforce it, so it seems very strange to me to even assert that this sort of thing would be the practical result of AI reimplementations. This seems to just be pointing towards the hypocrisy of one particular company, and assuming that this would be the inevitable universal, attitude, and result when there's no evidence to think so.
It's ironic, because antirez actually literally addresses this specific argument. They completely miss the fact that a lot of his blog post is not actually just about legal but also about ethical matters. Specifically, the idea he puts forward is that yes, corporations can do these kinds of rewrites now, but they always had the resources and manpower to do so anyway. What's different now is that individuals can do this kind of rewrites when they never have the ability to do so before, and the vector of such a rewrite can be from a permissive to copyleft or even from decompile the proprietary to permissive or copyleft. The fact that it hasn't been so far is a more a factor of the fact that most people really hate copyleft and find an annoying and it's been losing traction and developer mind share for decades, not that this tactic can't be used that way. I think that's actually one of the big points he's trying to make with his GNU comparison — not just that if it was legal for GNU to do it, then it's legal for you to do with AI, and not even just the fundamental libertarian ethical axiom (that I agree with for the most part) that it should remain legal to do such a rewrite in either direction because in terms of the fundamental axioms that we enforce with violence in our society, there should be a level playing field where we look at the action itself and not just whether we like or dislike the consequences, but specifically the fact that if GNU did it once with the ability to rewrite things, it can be done again, even in the same direction, it now even more easily using AI.
> They completely miss the fact that a lot of his blog post is not actually just about legal but also about ethical matters.
Honestly I was confused about the summarization of my blog post into just a legal matter as well. I hope my blog post will be able to flash at least a short time in the HN front page so that the actual arguments it contain will get a bit more exposure.
I'm failing to see what in the quoted text you took to be about AI rewrites specifically? It just reads as a slightly catty aside about the social reaction of rewrites in general (by implying the one example is generalizable.)
Perhaps we should finally admit that copyright has always been nonsense, and abolish this ridiculous measure once and for all
I think AI is very much eroding the legitimacy of copyright - at least to software, which is long been questioned since it's more like math than creative expression.
I think the industry will realize that it made a huge mistake by leaning on copyright for protection rather than on patents.
Probably a wiser approach is to consider different times require different measures (in general!).
I did not study in detail if copyright "has always been nonsense", but I do agree that nowadays some of the copyright regulations are nonsense (for example the very long duration of life + 70 years)
IMO the core idea of copyright isn't nonsense, but I do think the current implementation (70+ years after death) is egregiously overpowered. I've always thought the current laws were too deeply entrenched to ever change, but I'm tentatively optimistic AI will shock the system hard enough to trigger actual reform.
Actually I think the last 20 years of the Internet demonstrates that copyright is more important than ever, because unless it's enforced, people with more capital than the copyright owner will simply steal creative works and profit from them.
The idea that "information wants to be free" was always a lie, meant to transfer value from creators to platform owners. The result of that has been disastrous, and it's long past time to push the pendulum in the other direction.
I think we're going one step too far even, AI itself is a gray area and how can they guarantee it was trained legally or if it's even legal what they're doing and how can they assert that the input training data didn't contain any copyrighted data.
Google already spent billions of dollars and decades of lawyer hours proving it out as fair use. The legal challenges we see now are the dying convulsions of an already broken system of publishers and IP hoarders using every resource at their disposal to manipulate authors and creators and the public into thinking that there's any legitimacy or value underlying modern copyright law.
AI will destroy the current paradigm, completely and utterly, and there's nothing they can do to stop it. It's unclear if they can even slow it, and that's a good thing.
We will be forced to legislate a modern, digital oriented copyright system that's fair and compatible with AI. If producing any software becomes a matter of asking a machine to produce it - if things like AI native operating systems come about, where apps and media are generated on demand, with protocols as backbone, and each device is just generating its own scaffolding around the protocols - then nearly none of modern licensing, copyright, software patents, or IP conventions make any sense whatsoever.
You can't have horse and buggy traffic conventions for airplanes. We're moving in to a whole new paradigm, and maybe we can get legislation that actually benefits society and individuals, instead of propping up massive corporations and making lawyers rich.
Google has cut out some very specific ruling that have nothing to do with modern AI. These systems are just a really slow/lossy git clone, current law has no trouble with it, it's broadly illegal.
If corporations are allowed to launder someone else work as their own people will simply stop working and just start endlessly remixing a la popular music.