Meta sued by publishers over Llama

Books are the object here, but the fight is really about how an AI model gets built. On May 5, five major publishers and novelist Scott Turow sued Meta and Mark Zuckerberg in federal court in Manhattan, saying Llama was trained on massive amounts of copyrighted material taken without permission. The claim is not just that Meta read copyrighted works. It is that Meta knowingly pulled them from pirate sources and used them anyway. That makes this a much sharper case than the usual fuzzy “AI learned from the internet” argument. (publishers.org) ### Who is suing whom? The plaintiffs are Elsevier, Cengage, Hachette Book Group, Macmillan, McGraw Hill, and author Scott Turow. They filed a proposed class action against Meta and named Zuckerberg personally, saying he was directly involved in decisions around AI training and approved the use of pirated datasets for Llama. Meta(publishers.org) can be fair use. (publishers.org) ### What do they say Meta actually did? The core allegation is simple — Meta copied millions of books, journal articles, and other text works without paying for licenses. The complaint says Meta scraped copyrighted material from across th(publishers.org) web crawling. (publishers.org) ### Why does the 267-terabyte number matter? Because it turns an abstract complaint into something concrete. The suit says Meta used roughly 267 terabytes of pirated books, articles, and other material from sources including LibGen and Anna’s Archive. That detail matters because it suggests scale, intent, and a specific supply ch(publishers.org)legedly knew were unauthorized. (publishers.org) ### Why name Zuckerberg personally? That is the pressure point. Big copyright suits usually target the company. Here, the plaintiffs also say Zuckerberg “personally authorized and actively encouraged” the infringement. If that claim survives, it raises the stakes for discovery and makes the case feel less like a technical compliance dispute and more like a management-choice case. (cbsnews.com) ### Is this just about training data? No — the complaint also says Llama can generate summaries, stylistic imitations, and in some cases near-verbatim or verbatim reproductions of protected works. That matters because copyright cases get stronger when plaintiffs can point to outputs that look too much like the originals. T(cbsnews.com)ame lawsuit. (cbsnews.com) ### Why does this hit Llama differently? Because Llama is an open-weight model family. Meta does not just run it inside one app — it releases model weights that others can adapt and deploy. So the legal question is not only whether Meta owes damages. It is whether businesses building on top of Llama start worrying that ups(cbsnews.com)rget and from how open-weight models spread through the market. (publishers.org) ### What happens next? For now, this is a proposed class action, not a ruling. The court still has to deal with class certification, Meta’s defenses, and the basic question hanging over the whole AI industry — when does training on copyrighted material count as fair use, and when does it cross into plain copying? This case will n(publishers.org)ers. (publishers.org) ### Bottom line The big shift here is not just that Meta got sued again. It is that major publishers are trying to show a direct line from pirate libraries to a flagship AI model — and to the CEO who allegedly approved the shortcut. If they can prove that, the debate moves from “AI scraped too much” to “AI was built with stolen inventory.” (publishers.org)