It depends how you use the GPL code. Merely training an LLM with such code is arguably fair use. I personally disagree with that assessment, buy most courts have ruled it as fair use so far. To be honest, US copyright law just merely wasn’t written with modern LLMs in mind and it is desperately in need of a number of statutory updates.
The training, sure, that’s not what I’m talking about. I’m saying that any code output by a model trained on GPL code must be GPL, because it is derived from GPL code.
There is a little bit of confusion here I think. Copyright protects the expression itself (I.e., the code as it is written) and it allows the author (or owner) to stop people from infringing their code, which means performing, displaying, etc… The license – in your example, a GPL license – is just a contract that gives other people permission to do those things without committing copyright infringement. Fair use is a a defense someone can raise when they are being sued for copyright infringement.
The lawsuits against LLM companies are primarily based on the theory that LLM companies are infringing the copyright of the original authors beyond the scope of any license they might have, if any. The AI companies are saying that their scraping and use of the underlying work to train the model is fair use.
The thing you are talking about, the code being outputted by the LLMs, is sufficiently disconnected from the original copyrightable expression such that courts have so far ruled that any such output is not a derivative work of the original. This is mainly because such outputs have a randomness/temperature in how they are created, so it isn’t producing the same expression nor is it a derivative work under existing law.
There are a lot of other factors that go into all of this. For example, under Google v. Oracle, code generally gets “thinner” protection than literary or more creative works because it is functional in nature. There is an entire career’s worth of case law on this topic so it can get quite fact specific. But hopefully my explanation of the current system makes a little more sense.
That doesn’t align with the guidelines published by the U.S. copyright office. They claim outputs similar to training data do infringe on the copyright of the training data, effectively supporting the “derivative work” claim, since they do NOT have to be identical.
“The training process [also] implicates the right of reproduction.” – Part 3, III, B.
“outputs likely infringe the reproduction right and, to the extent they adapt the originals, the right to prepare derivative works. Some commenters noted that, depending on the content type and the audience, they may implicate the public display and public performance rights as well.” — Part 3, III, D
“The Office concludes that, given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output.” – Part 2, II, D, 2
“Affirming the Office’s refusal, the court stated that ‘copyright law protects only works of human creation’” – Part 2, II, B
“In most cases, however, humans will be involved in the creation process, and the work will be copyrightable to the extent that their contributions qualify as authorship.” – Part 2, II, B
Ah, well, thank you for the explanation. It’s not confusion - I disagree with the court decisions you’re describing.
Randomness or no, a copy-paste machine is still a copy-paste machine. It doesn’t do anything else. LLMs are language search algorithms with fancy optimization routines bolted on, nothing more. Everything else that’s been said about these “artificial intelligence” products is marketing.
Reaching a different conclusion about the “transformative nature” (or whatever) of LLM output requires understanding less about how they actually work.
Randomness or no, a copy-paste machine is still a copy-paste machine. It doesn’t do anything else. LLMs are language search algorithms with fancy optimization routines bolted on, nothing more.
This sort of extreme reductionism is not helpful. They are significantly different from a “copy-paste” machine. The original training data does not exist in the model itself. It’s very much a “transformative” process.
Reaching a different conclusion about the “transformative nature” (or whatever) of LLM output requires understanding less about how they actually work.
This sounds like you’re saying “wilful ignorance is better than knowledge” - surely that’s not what you meant though?
The ideal is that the training data don’t exist on it. The reality is that the training algorithm is a very very weird compression algorithm and sometimes the model contains just derived properties and sometimes you can perform sample recovery and get entire books back out. That which is recoverable with a prompt less complex than the output is inherently already in the model, because that’s simply how entropy works.
It depends how you use the GPL code. Merely training an LLM with such code is arguably fair use. I personally disagree with that assessment, buy most courts have ruled it as fair use so far. To be honest, US copyright law just merely wasn’t written with modern LLMs in mind and it is desperately in need of a number of statutory updates.
The training, sure, that’s not what I’m talking about. I’m saying that any code output by a model trained on GPL code must be GPL, because it is derived from GPL code.
There is a little bit of confusion here I think. Copyright protects the expression itself (I.e., the code as it is written) and it allows the author (or owner) to stop people from infringing their code, which means performing, displaying, etc… The license – in your example, a GPL license – is just a contract that gives other people permission to do those things without committing copyright infringement. Fair use is a a defense someone can raise when they are being sued for copyright infringement.
The lawsuits against LLM companies are primarily based on the theory that LLM companies are infringing the copyright of the original authors beyond the scope of any license they might have, if any. The AI companies are saying that their scraping and use of the underlying work to train the model is fair use.
The thing you are talking about, the code being outputted by the LLMs, is sufficiently disconnected from the original copyrightable expression such that courts have so far ruled that any such output is not a derivative work of the original. This is mainly because such outputs have a randomness/temperature in how they are created, so it isn’t producing the same expression nor is it a derivative work under existing law.
There are a lot of other factors that go into all of this. For example, under Google v. Oracle, code generally gets “thinner” protection than literary or more creative works because it is functional in nature. There is an entire career’s worth of case law on this topic so it can get quite fact specific. But hopefully my explanation of the current system makes a little more sense.
That doesn’t align with the guidelines published by the U.S. copyright office. They claim outputs similar to training data do infringe on the copyright of the training data, effectively supporting the “derivative work” claim, since they do NOT have to be identical.
“The training process [also] implicates the right of reproduction.” – Part 3, III, B.
“outputs likely infringe the reproduction right and, to the extent they adapt the originals, the right to prepare derivative works. Some commenters noted that, depending on the content type and the audience, they may implicate the public display and public performance rights as well.” — Part 3, III, D
“The Office concludes that, given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output.” – Part 2, II, D, 2
“Affirming the Office’s refusal, the court stated that ‘copyright law protects only works of human creation’” – Part 2, II, B
“In most cases, however, humans will be involved in the creation process, and the work will be copyrightable to the extent that their contributions qualify as authorship.” – Part 2, II, B
All quotes from: https://www.copyright.gov/ai/
Especially image models where you can literally just ask for and get images from copyright protected works like art, movies, comics, etc
Ah, well, thank you for the explanation. It’s not confusion - I disagree with the court decisions you’re describing.
Randomness or no, a copy-paste machine is still a copy-paste machine. It doesn’t do anything else. LLMs are language search algorithms with fancy optimization routines bolted on, nothing more. Everything else that’s been said about these “artificial intelligence” products is marketing.
Reaching a different conclusion about the “transformative nature” (or whatever) of LLM output requires understanding less about how they actually work.
This sort of extreme reductionism is not helpful. They are significantly different from a “copy-paste” machine. The original training data does not exist in the model itself. It’s very much a “transformative” process.
This sounds like you’re saying “wilful ignorance is better than knowledge” - surely that’s not what you meant though?
The ideal is that the training data don’t exist on it. The reality is that the training algorithm is a very very weird compression algorithm and sometimes the model contains just derived properties and sometimes you can perform sample recovery and get entire books back out. That which is recoverable with a prompt less complex than the output is inherently already in the model, because that’s simply how entropy works.