The concern related to copyright is not that the AI labs would assert copyright over the code produced by their LLMs - that is a complete strawman.
The concern instead is that LLMs and all of their outputs may be found to be derivative works of their entire training set, and thus rendered unusable (as the training set is not distirbutable under any license).
I think this ship has long sailed and no court is going to dare give such a decision given the money involved, for better or for worse. But it's a much more realistic scenario, in principle, than LLM labs going mad and attacking their own customers.
Edit to add: there is another, completely different, copyright risk associated with LLMs - and one that is much more realistic. It is the fact that code generated by LLMs may not, in fact, be copyrightable at all. Which would mean that it can't be subject to the GPL. As long as it remains a minority of GCC code, this wouldn't matter much, but it could in time lead to significant portions of GCC becoming public domain, and thus cooyable, modifiable, and redistrubutable without providing the four freedoms.
The mental model of slow decay into copyable is complicated. The code base is already a wash with contributions that have low copyrightability in US case law anyway, things derived directly from external architecture or documentation in barely novel ways, common system interfaces, and so on. LLM output is another output of this kind of class and sure sometimes large contributions happen that seem more significant, but they likely have little bearing on the work as a whole given it's earlier design provenance (at least for established and massive projects like this).
What is likely to get more muddy over time is the accuracy of any copyright registration, and the enforcement of copyright infringements on portions of the whole. These are already complicated cases and definitely so for compilers with so much "scènes à faire".
It's not clear how much this has a negative impact on cases around the whole, which tend to be the more important cases for the four freedoms that, while they have other intentions, have a primary intention of ensuring that the whole continues to be available for redistribution and extension in perpetuity.
I do not think that there is a clear link between these two areas at all, and the GPL's most important intents may be far safer long term than concerns of dilution suggest.
Courts have already set a precedent on this (it's fair use) and even if these decisions were overturned, the liability would fall on the LLM creators for the initial training, not on their users. The outputs would keep being treated individually on their own merits.
I believe there are still ongoing lawsuits on this in various places, I don't believe it's fully settled law - though, as I mentioned in my post, I also don't expect there is any realistic chance for the courts to simply shut down a trillion dollar industry, as such a decision likely would.
I do believe though that, if the LLMs were found to be derivative works of their training set, it would follow almost directly that their output is also a derivative work of that same training set - given how these LLMs operate. And even if the liability fell with the LLM providers (which may not be so clear cut for, say, local models, fine tuning, etc), that would still mean everyone would have to excise any LLM generated content they are distributing.
> were found to be derivative works of their training set, it would follow almost directly that their output is also a derivative work
I doubt so. Let's say Harry Potter is in the training set and you ask the LLM to generate a quick sort function in C, is that quick sort function a derivative of Harry Potter? What if you ask the LLM to output some known public domain work? That leads to a contradiction where according to one definition, the work is public domain and according to the other, it is a derivative of Harry Potter. It seems to me that there's no other option but to consider each output on its own merit.
Not really, no. There have been many settlements of lawsuits, which means they are very explicitly not precedents. The LLM companies likely prefer to settle because of the existential risk being found liable for copyright infringement would result in, whereas even an expensive settlement doesn’t upend their entire business model.
Furthermore, I don’t think you can really assume that the courts will rule a certain way on this just because of the money involved; there’s a lot of money involved when it comes to the copyright holders too, and they’ve long enjoyed a rather favorable status with the courts and legislators. (For example, in the days of P2P file sharing lawsuits and attempts to legislate P2P file sharing, the software industry was already many times the size of the media industry, but the media industry consistently won.)
> Furthermore, I don’t think you can really assume that the courts will rule a certain way on this just because of the money involved; there’s a lot of money involved when it comes to the copyright holders too, and they’ve long enjoyed a rather favorable status with the courts and legislators.
I don't think this is all that plausible, even though I agree with you that it's not settled law. The size of the AI industry is gigantic, and a ruling that they are infringing the copyright of every piece of content in their training set would essentially shut them down entirely. Such a decision, if final, would probably easily wipe out a few hundred billion dollars on the stock market. Even if any court was willing to go that far, almost certainly lawmakers would step in and modify copyright law to prevent this from happening - both in the USA and the EU.
I don't think there is any comparison to make with the file sharing battle. That was a much, much smaller industry, it was not a significant chunk of the total hardware and software industries. Plus, the software titans were not nearly as well connected politically as they are today.
A couple of things: The “AI industry” is nowhere near as large as it pretends to be, most of its valuation is in the form of IOUs being traded back and forth. However, the bubble popping probably will wipe a few hundred billion of value from the stock market, because that’s what happens when a speculative bubble pops.
The second thing is that I’m not necessarily talking about whether _a specific LLM itself_ infringes copyright, but whether _its output_ is covered by the copyright of _its training material_. Whether training an LLM is an activity that infringes copyright is not well-settled in any precedential way, whether the trained LLM as an artifact infringes copyright is even less settled, and whether the output of that LLM is either infringing or covered by copyright is also not settled. These are all still extremely open questions.
That means anyone doing reasonable risk management should not just blithely race ahead and assume that there’s no infringement, which appears to be the approach the GCC project is taking explicitly and which also appears to be the approach projects like Linux and LLVM are taking implicitly (mostly through weasel-language like accepting responsibility for code you’re submitting).
"To summarize the analysis that now follows, the use of the books at issue to train Claude and its precursors was exceedingly transformative and was a fair use under Section 107 of the Copyright Act. And, the digitization of the books purchased in print form by Anthropic was also a fair use but not for the same reason as applies to the training copies. Instead, it was a fair use because all Anthropic did was replace the print copies it had purchased for its central library with more convenient space-saving and searchable digital copies for its central library — without adding new copies, creating new works, or redistributing existing copies."
IANAL and don't know how significant this decision is, but it is, at the very least, how one judge views it.
Personally, I don't think judges will rule a certain way because of the money involved but because it seems clear that training a ML model is highly transformative.
Not quite the precedent that it may sound like. A district court judge ruled that using copyrighted materials for the training itself was not infringement, but that the materials must be obtained legally.
Anthropic is trying to settle the case with most plaintiffs with respect to obtaining their works in an infringing way, but there are still plaintiffs pursuing the case on both the grounds that the remedy is insufficient (being only about $3000/work, when it has been as high as $250K/work in other copyright infringement cases and via statutory damages) and also on the grounds that the ruling that training is fair use was an error of law on the district court judge’s part.
Notably it doesn’t cover whether the output of the trained LLM continues to attach the training set’s copyright, which is independent from whether the training itself was an infringing activity. And there’s a substantial argument that the judge erred, if it can be shown that the training works are stored in a recoverable manner (even with some loss/defredation) rather than more extensively transformed.
In many timelines, non-copyrightable code in GPL project doesn’t restrict the user’s freedom, since an LLM can easily reimplement the functionality clean-room. AI has brought levels of freedom to users that GNU could never do. Practical levels of freedom, leading to real, positive outcomes. Software is finally free in the real world.
For certain use cases, probably. For others, no. For example, hardware that uses GNU software must be patchable by users (if it is patchable at all). Hardware running public software can be completely locked down to its users, regardless of how powerful of an LLM the user has access to.
The concern instead is that LLMs and all of their outputs may be found to be derivative works of their entire training set, and thus rendered unusable (as the training set is not distirbutable under any license).
I think this ship has long sailed and no court is going to dare give such a decision given the money involved, for better or for worse. But it's a much more realistic scenario, in principle, than LLM labs going mad and attacking their own customers.
Edit to add: there is another, completely different, copyright risk associated with LLMs - and one that is much more realistic. It is the fact that code generated by LLMs may not, in fact, be copyrightable at all. Which would mean that it can't be subject to the GPL. As long as it remains a minority of GCC code, this wouldn't matter much, but it could in time lead to significant portions of GCC becoming public domain, and thus cooyable, modifiable, and redistrubutable without providing the four freedoms.