TL;DR: The advent of AI based, LLM coding applications like Anthropic’s Claude and ChatGPT have prompted maintainers to experiment with integrating LLM contributions into open source codebases.
This is a fast path to open source irrelevancy, since the US copyright office has deemed LLM outputs to be uncopyrightable. This means that as more uncopyrightable LLM outputs are integrated into nominally open source codebases, value leaks out of the project, since the open source licences are not operative on public domain code.
That means that the public domain, AI generated code can be reused without attribution, and in the case of copyleft licences - can even be used in closed source projects.
Ok, but it’s not like everyone is documenting exactly which parts are generated, curated, or human written.
Say someone incorporates GPL code without attribution, and gets sued for doing so. They try to make the argument in court that the source material they used is not copyrighted, because of AI. Won’t they have to prove that the parts they used were actually AI output for this defense to work? It isn’t like people are going around ignoring the copyright on things in general if they look like they were probably generated with AI, that isn’t enough to be safe from prosecution, because you usually can’t know the exact breakdown. It seems like preventing this loophole from being used would be as simple as keeping it ambiguous and not allowing submissions that positively affirm being entirely AI generated.
I don’t really think we need to go down the copyfraud path to see that AI code damages copyleft projects no matter what - we know that some projects are already accepting AI generated code, and they don’t ask you to hide it - it is all in the open.
I don’t see how that follows or contradicts what I’m saying though. They could hide it, easily. Even if they don’t hide it, how useful would it really ever be to only use the portions of the codebase that have been labelled as having been AI generated? Can one even rely on those labels? Making use of the non-copyrightability of AI output to copy code in otherwise unauthorized ways does not seem like a straightforward or legally safe thing to do. That’s especially the case because high profile proprietary software projects also make heavy use of AI, it doesn’t seem likely the courts will support a legal precedent that strips those projects of copyright and allow anyone to use them for whatever. So basically I’m not at all convinced about the idea that AI code damages copyleft projects, it seems unlikely to be a problem in practice.
I think what may happen in practice could be worse - basically if we can’t tell whether some code is the work of a human, but the project accepts AI code, if there we forego the analysis of whether something was produced by a human, the entire project may be deemed public domain – perhaps after a certain date (when LLM contributions were welcomed).
Beyond that, by integrating LLM code into those projects, the projects are signifying assent to their works to be consumed by LLMs for infringement of the whole work - not just the LLM produced portions - it is hard to be doctrinaire about adherence to the open source license when the maintainers themselves are violating it.
We may see a future where copyrights for works become more like trademarks - if you don’t make any attempt to protect your work from piracy, you may simply lose the right to contest its theft.
Obviously, it is as you say - today the courts may smile upon a GPL project where a commercial vendor copied and released as their own without sharing alike - but if they instead say that they copied the work into their LLM and produced a copy without protections (as chardet has done), the courts might be less willing to afford the project copyright protections if the project itself was making use of the same copyright stripping technology to strip others’ work to claim protections over copied work.
Besides which, “authored by Claude” seems like a pretty easy way to find public domain code, and as Malus presents, the only code that may ultimately be protected is closed source code - you can’t copy it if you don’t have the source.
The diversion of “people may try to pass of LLM code as their own” is a nice diversion, but ancillary to the existing situation where projects are incorporating public domain code as licensed. We can start there before we start worrying about fraud.
ianal but does it even work like that? Is there any specific reason to think it does? I don’t believe you really get credit for purity and fairness vibes in the legal system. Same goes for the idea that code where it is ambiguous whether it is AI output could be considered public domain, seems kind of implausible, is there actually any reason to think the law works that way? If it did, then any copyrighted work not accompanied by proof of human authorship would be at risk, uncharacteristic for a system focused on giving big copyright holders what they want without trouble.
There is no way, leaks happen, big tech companies have massive influence, a situation where their code falls into the public domain as soon as the public gets their hands on it just isn’t realistic. I feel suspicious that many of these concerns are coming from a place of not wanting LLM code in open source projects for other reasons, rather than the existence of a strong legal case that it represents a real and serious threat to copyleft licensing.
I’m mostly just playing along with your thought experiment. As I said, we know that projects are already accepting LLM code into projects that are nominally copyleft.
If that is the case, is chardet 7.0.0 a derivative work of chardet, or is it a public domain LLM work? The whole LLM project is fraught with questions like these, but it seems that the vendors at least are counting on not copying leaked software and instead copying open source code that is publicly hosted.
Why is it okay to strip copyright from open source works but not from leaked closed source works?
We know that Disney is suing to protect its works - if it is true that LLM outputs are transformative, they should lose, as should any vendor whose leaked code was “transformed” by an LLM.
I think the reimplementation stuff is a separate question because the argument for it working looks a lot stronger, and because it doesn’t have anything to do with the source material having LLM output in it. Also if this method holds as legally valid, it’s going to be easier to just do that than justify copying code directly (which would probably have to only be copies of the explicitly generated parts of the code, requiring figuring out how to replace the rest), which means it won’t matter whether some portion of it was generated. I don’t see much reason to think that a purist approach to accepting LLM code will offer any meaningful protection.
So what though? If they aren’t entirely generated, you can’t make a full fork, and why would a partial fork be useful? If it isn’t disclosed what parts are AI, you can’t even do that without risking breaking the law.
Is it a separate question, though?
Both works are copyrighted, one is just copyrighted as “all rights reserved” (our leaked commercial code) and the rest is licensed as LGPL. We’re putting both pieces of code inside the LLM and then asking the LLM to make a new version.
What makes the action of leaking different from the act of putting it on the web? Rights are reserved in either case.
Well, people are contributing to copyleft codebases expecting that when people build on their work, that work (the derivative works) are also licensed in the same way. You don’t need to fork for the value to be lost. People expected virality to be part of their contribution, and clearly the new derivative works are partially non-copyleft.
Beyond that, as more of the codebase is LLM produced, the less of it is protected by the copyleft license, until we have a ship of Theseus situation where the codebase is available, but no longer copyleft. That is clearly not what was intended by e.g. the GPL. Just look at the Stallman quote in post.