Five publishers sue Meta over Llama training data
Five major publishers and author Scott Turow allege Meta used millions of copyrighted books and journal articles to train its Llama AI models.
Five major publishers and author Scott Turow have filed a proposed class-action lawsuit against Meta Platforms and chief executive Mark Zuckerberg, alleging that millions of copyrighted books and academic articles were copied without permission to train the company’s Llama artificial-intelligence models.
Elsevier, Cengage Learning, Hachette Book Group, Macmillan Publishers and McGraw Hill filed the complaint in the US District Court for the Southern District of New York on May 5. Turow and his company, S.C.R.I.B.E. Inc, are also named as plaintiffs.
The plaintiffs are seeking damages and an injunction against the alleged infringement. Their requested remedies include the destruction of unauthorised copies of protected works obtained or retained by Meta. The case remains a proposed class action and has not yet been certified by the court.
Meta has maintained that the use of copyrighted material to train artificial-intelligence systems can be protected under the US fair-use doctrine. The company has pledged to contest the latest lawsuit vigorously, pointing to previous court decisions that treated some forms of AI training as transformative use.
Publishers allege use of pirate libraries
The complaint alleges that Meta obtained copyrighted material from online pirate libraries and other unauthorised sources, including through torrent downloads, before using the works to develop Llama.
The publishers also accuse Meta of removing or disregarding copyright-management information associated with the material and proceeding without seeking licences or negotiating with rights holders. They argue that Meta used the works to build commercially valuable products that could compete with authors and publishers.
Zuckerberg is named personally in the complaint. The plaintiffs allege that he knew about and authorised decisions concerning the acquisition and use of copyrighted training material. Those claims have not been decided by the court.
The case is one of a growing number brought by authors, publishers, news organisations, artists and other creators against companies developing generative-AI systems. The central disputes concern whether copying works for model training constitutes fair use and whether companies can rely on that defence when the material was acquired from unauthorised sources.
Training datasets expand rapidly
The scale of information used to train leading AI models has increased sharply as technology companies compete to improve their systems.
Meta reported that Llama 3.1 was trained on more than 15 trillion tokens. Model documentation for Llama 4, released in April 2025, stated that its Scout and Maverick models were pre-trained on as many as 40 trillion tokens.
Chinese developer DeepSeek used 8.1 trillion tokens to train DeepSeek-V2 before increasing the total to 14.8 trillion for DeepSeek-V3. Epoch AI’s model database similarly records a broad rise in training-data volumes across large language models.
Token totals do not represent the number of separate books, articles or other documents used, nor do they establish whether the underlying material was acquired lawfully. They nevertheless illustrate the scale of the datasets now required to develop leading AI systems.
Courts distinguish training from data acquisition
Previous rulings have offered technology companies some protection for using copyrighted works in AI training, but courts have also drawn distinctions based on how those works were obtained and stored.
In June 2025, US District Judge Vince Chhabria ruled in Meta’s favour on fair-use claims brought by a group of authors that included comedian Sarah Silverman and writers Richard Kadrey and Christopher Golden.
The judge found that the plaintiffs had not presented sufficient evidence of market harm on the record before the court. However, he stressed that the decision was narrow and did not establish that training AI systems with copyrighted material was lawful in every case. Other issues in the litigation have continued.
A separate lawsuit against Anthropic, the developer of Claude, produced a similar distinction. A federal judge ruled that using books to train a large language model was transformative and could qualify as fair use, but found that retaining millions of pirated copies in a permanent central library was not protected by the same doctrine.
Anthropic later agreed to a US$1.5 billion settlement covering claims related to the pirated copies. A federal judge granted final approval to the settlement on July 20, 2026.
The lawsuit against Meta remains pending in Manhattan federal court. The court has yet to rule on the publishers’ allegations, Meta’s fair-use defence or whether the proposed class action may proceed on behalf of a wider group of rights holders.
