Meta is currently embroiled in several lawsuits in California federal court regarding the use of pirated books to train its Llama AI models. While a judge previously ruled that training on books could constitute fair use, the specific method of acquiring those books remains a primary point of contention. The litigation involves high-profile authors like Sarah Silverman and Ta-Nehisi Coates, as well as major publishers such as Entrepreneur Media and Chicken Soup for the Soul.
Recent deposition testimony from Stephen Roller, a former Meta researcher who worked in the Fundamental AI Research (FAIR) group between 2017 and 2023, has provided new details for the plaintiffs. According to a joint status report, Roller admitted to torrenting the Books3 dataset—a collection of nearly 200,000 books sourced from the private torrent tracker Bibliotik—and knowingly distributing it. Roller testified that Meta had copies of the dataset by the time he left the company and identified two other research scientists who likely acquired their own copies. He also confirmed that he deleted his personal copy of the dataset about a year ago, citing the ongoing litigation as the reason for its removal.
This image relates to the specific testimony of Stephen Roller regarding the torrenting and distribution of the Books3 dataset.
Meta has contested the plaintiffs' interpretation of this testimony. The company argues that Roller did not specify that his download was performed for Meta's purposes or on Meta's infrastructure, suggesting it could have been for personal use or a different employer. Furthermore, Meta maintains that it has already produced substantial quantities of server logs and is continuing to provide files on a rolling basis. The plaintiffs, however, continue to press for more specific data, including command history files from torrenting servers and the contents of an internal Google Drive folder titled "ML Books," which they believe may contain evidence of pirated material used for non-AI purposes, such as a company book club. The Entrepreneur Media case is currently scheduled for trial in the spring of 2027, with several other publishers joining the litigation queue in the interim.
This image illustrates Meta's defense regarding the researcher's actions and the company's production of server logs.