OpenAI's justification for why training data is fair use, not infringement [pdf]

Re:OpenAI's justification for why training data is fair use,... by pclmulqdq on Friday October 06, 2023 @01:22PM (#37793491)

I would assume that #2 is actually "yes" given the way derived works work and the fair use tests, and that #1 and #2 are actually very much linked. Fair use is a defense to copyright infringement, and it's a relatively complex balance of factors. It's relatively inarguable (even OpenAI isn't arguing this) that the model isn't a derivative work of its training set: they are just arguing that they have fair use rights to the contents of the training set. Only one of the factors in a fair use analysis is how transformative the use is, and I think it's hard to argue that training an LLM isn't a huge transformation. However, the other factors weigh pretty heavily against LLMs here, and the Author's Guild lawsuit is a pretty good set of arguments as to why. It's up to a court to decide whether the transformative nature outweighs the other factors.

If you're lumping the fair use question into #1 and the "is it a derived work" question is #2, I'm pretty sure that nobody on any side of this agrees that it isn't a derived work. Once you have a derived work, you can either ask whether they were licensed to produce that work (no) or whether it was fair use (possibly).

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by lsaferite on Friday October 06, 2023 @05:10PM (#37796459)

On a purely technical level, I feel very strongly that the model is not a derivative work any more than a dictionary is a derivative work. The model is a recording of token stream probabilities. The training data is, in my mind, where the conversation about derivative work and fair use comes into play. Again, I'm not a lawyer and have little exposure to the legal profession, so my opinions are my own.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by pclmulqdq on Saturday October 07, 2023 @10:05AM (#37801800)

There's a big difference between a model and almost every other form of compilation of human knowledge, which is when the human intervention occurs. A dictionary is a collection of definitions from humans that is derived from their experience reading texts. Human knowledge and experience moderates the flow from the text the human reads to the text in the dictionary. LLM training has no such moderation, and that is a significant factor in whether something is a derived work or not.

This is really all up to the courts, and I don't think anyone is confident about how it will shake out. However, the fact that OpenAI isn't even trying to make the argument that an LLM isn't a derived work of the training set (they are going straight to fair use, which is an acknowledgement of infringement) suggests that this is not actually contentious.

[ Parent | on HN ]