It's reasonably likely, but not yet settled, that LLM training falls under fair use and doesn't require a license. This is what the https://githubcopilotlitigation.com/ class action (from 2022) is about, and its still making its way through the court.
This prediction market has it at 12% likely to succeed, suggesting that courts will not agree with you: https://manifold.markets/JeffKaufman/will-the-github-copilot...
It's reasonably likely, but not yet settled, that LLM training falls under fair use and doesn't require a license.
I would say it's not reasonably likely that LLM training is fair use. Because I've read the most recent SCOTUS decision on fair use (Warhol), and enough other decisions on fair use, to understand that the primary (and nearly only, in practice) factor is the effect on the market for the original. And AI companies seem to be going out of their way to emphasize that LLM training is only going to destroy the market for the originals, which weighs against fair use. Not to mention the existence of deals licensing content for LLM training which... basically concedes the point.
Of the various options, a ruling that LLM training is fair use I find the least likely. More likely is either that LLM training is not fair use, that LLM training is not infringing in the first place, or that the plaintiffs can't prove that the LLM infringed their work.
I do not read it that way at all. The Goldsmith decision mainly turns on the idea that an artist protections include that for derivative works. Warhol produced a work that does substantially the same things as Goldsmith's, ie, is a picture that can be viewed.
When talking about parody, they note that the usage as the foundation for parody is always substantially different from the original and thereby allowed, even if it would otherwise infringe. LLMs are always substantially different from the original, too.
If I want to write software that draws that picture exactly, the code would not be a copyright violation. It is text and cannot be printed in a magazine as a picture. If I used it to print a picture that was a derivative work and sold that, it might be.
A large language model has no intersection with the picture or, for that matter, anything that it absorbs. It is possible that someone might figure out how to prompt it to do exactly the same picture as Goldsmith did but fairly unlikely.
Unless you could show that this was easy, common and part of the intent of the LLM creator, I can see no possibility that it is infringing.
This prediction market has it at 12% likely to succeed
Randos on the internet with a betting addiction are distinctively different from a court of law. I wish people would stop talking about prediction market as if they mattered.
this isn't about copyright but about computer access. the CFAA is extremely broad; if you ban LLM companies from access on grounds of purpose you have every legal right to do so
in theory that legislation has teeth, too. they are not allowed to access your system if you say they are not; authentication is irrelevant.
every GET request to a system that doesn't permit access for training data is a felony
Comments
It's reasonably likely, but not yet settled, that LLM training falls under fair use and doesn't require a license. This is what the https://githubcopilotlitigation.com/ class action (from 2022) is about, and its still making its way through the court. This prediction market has it at 12% likely to succeed, suggesting that courts will not agree with you: https://manifold.markets/JeffKaufman/will-the-github-copilot...
I would say it's not reasonably likely that LLM training is fair use. Because I've read the most recent SCOTUS decision on fair use (Warhol), and enough other decisions on fair use, to understand that the primary (and nearly only, in practice) factor is the effect on the market for the original. And AI companies seem to be going out of their way to emphasize that LLM training is only going to destroy the market for the originals, which weighs against fair use. Not to mention the existence of deals licensing content for LLM training which... basically concedes the point.
Of the various options, a ruling that LLM training is fair use I find the least likely. More likely is either that LLM training is not fair use, that LLM training is not infringing in the first place, or that the plaintiffs can't prove that the LLM infringed their work.
I do not read it that way at all. The Goldsmith decision mainly turns on the idea that an artist protections include that for derivative works. Warhol produced a work that does substantially the same things as Goldsmith's, ie, is a picture that can be viewed.
When talking about parody, they note that the usage as the foundation for parody is always substantially different from the original and thereby allowed, even if it would otherwise infringe. LLMs are always substantially different from the original, too.
If I want to write software that draws that picture exactly, the code would not be a copyright violation. It is text and cannot be printed in a magazine as a picture. If I used it to print a picture that was a derivative work and sold that, it might be.
A large language model has no intersection with the picture or, for that matter, anything that it absorbs. It is possible that someone might figure out how to prompt it to do exactly the same picture as Goldsmith did but fairly unlikely.
Unless you could show that this was easy, common and part of the intent of the LLM creator, I can see no possibility that it is infringing.
Randos on the internet with a betting addiction are distinctively different from a court of law. I wish people would stop talking about prediction market as if they mattered.
Participants in prediction market do not need to be experts for their collective input to be informative.
There's a long history of economic research on the "wisdom of crowds" that backs up their value.
this isn't about copyright but about computer access. the CFAA is extremely broad; if you ban LLM companies from access on grounds of purpose you have every legal right to do so
in theory that legislation has teeth, too. they are not allowed to access your system if you say they are not; authentication is irrelevant.
every GET request to a system that doesn't permit access for training data is a felony
why are we pretending that these gambling sites have any weight on anything
What do you mean by weights?
I'd certainly trust their predictions more than those given by most "experts".