In order to help train its AI models, Meta (and others) have been using pirated versions of copyrighted books, without the consent of authors or publishers. The company behind Facebook and Instagram faces an ongoing class-action lawsuit brought by authors including Richard Kadrey, Sarah Silverman, and Christopher Golden, and one in which it has already scored a major (and surprising) victory: The Californian court concluded last year that using pirated books to train its Llama LLM did qualify as fair use.
You’d think this case would be as open-and-shut as it gets, but never underestimate an army of high-priced lawyers. Meta has now come up with the striking defense that uploading pirated books to strangers via BitTorrent qualifies as fair use. It further goes on to claim that this is double good, because it has helped establish the United States’ leading position in the AI field.
Meta further argues that every author involved in the class-action has admitted they are unaware of any Llama LLM output that directly reproduces content from their books. It says if the authors cannot provide evidence of such infringing output or damage to sales, then this lawsuit is not about protecting their books but arguing against the training process itself (which the court has ruled is fair use).
Judge Vince Chhabria now has to decide whether to allow this defense, a decision that will have consequences for not only this but many other AI lawsuits involving things like shadow libraries. The BitTorrent uploading and distribution claims are the last element of this particular lawsuit, which has been rumbling on for three years now, to be settled.
Arguing that training models isn’t fair use us going to be a massive uphill battle, it’s basically reading the book but with a computer. It’s not actually a big deal to people, unless you hold the copyright to a ton of works and want to get a percentage of all the AI income these companies have made.
Torrenting the books is likely absolutely copyright infringement, but that has relatively low payout compared to the money these companies are getting for their models. The training being fair use means that rights holders can’t try to take any money from the model’s use. The statutory limits for infringement even at per work levels aren’t significant compared to the legal cost of proving it happened.
Anthropic pirating books for their training corpus resulted in the biggest copyright settlement in history–well over a billion. That is still being quibbled over i believe, but they settled because they were likely to pay out more if the case went forward. So I’m not really sure where you’re coming from that infringement via torrenting does not result in monstrously large liability.
The judge in that case ruled the training wasn’t fair use for pirated books, which left them on the hook for potentially all revenue (likely a court determined percentage) that the model generated for them in addition to statutory damages. That is well north of 1.5 billion.
Which is kind of a pity. Anyone who’s ever written something on the net should be getting royalty checks from these fucks. I’m not exactly famous but I’ve written prolifically in my field of work and have gotten nearly word-for-word reproductions of my articles out of every big model I’ve tested since GPT-3.
There’s an argument to be made that it is, in fact, not ‘reading’. The training of the model could be considered a lossy compression of the data. And streaming movies in a lossy compression format is not fair use, is it?
The model doesn’t stream out anyone’s content though. The article mentions that the plaintiffs have provided no examples of a prompt that creates anything substantial.
Streaming a lossy compression would generally be infringement, but there is definitely a point where it becomes not infringement if it’s lossy enough.
What a model generally stores, is factual information that isn’t copyright in the first place. It’s storing word counts, sentence lengths, sentiment analysis, and so on.
It’s not the storage of the information that matters as much as the presentation. Google’s search index stores a huge amount of copyrighted material, even losslessly. But they only present small snippets at a time which is not considered copyright infringement. The question really is whether or not the information being presented by the models is in a format which is considered copyright infringement. So far, courts have not found that they are.