> The question is, if they could and would have paid for each book, would it be ok to train the LLM on them?
Whether training on AI model on an array of diffentent works, many of which are copyright protected, is itself a copyright violation, in addition to or distinct from any copyright violation that goes on gathering the dataset for training (and separate from any copyright violation in the actual or intended use of the LLM), remains to be resolved as a legal question, and may or may not have a simple yes or no answer (or the same answer under every system of copyright laws globally).
My inclination is that it is probably generally not a violation in US law, but that's not something I am very confident in; how the definitions of copy and derivative work apply to determine if it would be without fair use, and how fair use analysis applies, are not clear from the available precedent.
> But legally, how does using a book to train a LLM differ from a teacher learning from a book and teaching its contents to their pupils.
It is very clear, by looking at how US copyright law is written and even more clear in its history of application, that information stored in brains of people are without exception neither copies nor new works that can be derivative works under US law, and so cannot be infringing, no matter how you gain them. It’s also very clear in the statute itself and the case law that data in media used by artificial digital computers, on the other hand, can constitute copies or derivative works that can be infringing. Even if the process is arguably similar in legally relevant manners, copyright law is critically focussed on the result and whether it is a particular kind of thing which can be infringing, not just the process.
Whether training on AI model on an array of diffentent works, many of which are copyright protected, is itself a copyright violation, in addition to or distinct from any copyright violation that goes on gathering the dataset for training (and separate from any copyright violation in the actual or intended use of the LLM), remains to be resolved as a legal question, and may or may not have a simple yes or no answer (or the same answer under every system of copyright laws globally).
My inclination is that it is probably generally not a violation in US law, but that's not something I am very confident in; how the definitions of copy and derivative work apply to determine if it would be without fair use, and how fair use analysis applies, are not clear from the available precedent.
> But legally, how does using a book to train a LLM differ from a teacher learning from a book and teaching its contents to their pupils.
It is very clear, by looking at how US copyright law is written and even more clear in its history of application, that information stored in brains of people are without exception neither copies nor new works that can be derivative works under US law, and so cannot be infringing, no matter how you gain them. It’s also very clear in the statute itself and the case law that data in media used by artificial digital computers, on the other hand, can constitute copies or derivative works that can be infringing. Even if the process is arguably similar in legally relevant manners, copyright law is critically focussed on the result and whether it is a particular kind of thing which can be infringing, not just the process.