Recently unsealed court documents have shed light on internal discussions within Meta about using copyrighted works to train the company's AI models, despite potential legal implications. The documents, submitted as part of the Kadrey v. Meta case, reveal that Meta employees, including CEO Mark Zuckerberg, were aware of the legal risks but still considered using copyrighted materials, such as books, to improve their AI models.

The plaintiffs in the case, including authors Sarah Silverman and Ta-Nehisi Coates, argue that Meta's use of copyrighted works without permission constitutes copyright infringement. Meta, on the other hand, claims that training models on IP-protected works is "fair use." The new filings provide insight into the company's internal discussions and decision-making processes, which may have significant implications for the case.

In one internal chat, Meta research engineer Xavier Martinet suggested acquiring books and escalating the decision to executives, rather than seeking permission from publishers. Martinet argued that many startups were likely already using pirated books for training and that Meta could take a similar approach. His colleague, Melanie Kambadur, senior manager for Meta's Llama model research team, cautioned that using publicly available data would require approvals, but noted that Meta's lawyers were being less conservative in their approvals.

The documents also reveal that Meta considered using Libgen, a links aggregator that provides access to copyrighted works, as an alternative to licensed data sources. Libgen has been sued multiple times for copyright infringement and has been ordered to shut down. Despite this, some Meta employees believed that using Libgen was essential to staying competitive in the AI race. In an email, Sony Theakanath, director of product management at Meta, outlined "mitigations" to reduce Meta's legal exposure, including removing data marked as pirated and not publicly citing Libgen's use.

Other revelations in the filings suggest that Meta may have scraped Reddit data for model training and considered overriding past decisions on training data to ensure the company's models had sufficient data. The plaintiffs have amended their complaint several times since the case was filed in 2023, alleging that Meta cross-referenced pirated books with copyrighted books available for license to determine whether to pursue a licensing agreement with a publisher.

The case has significant implications for the AI industry, as it raises questions about the use of copyrighted materials in AI model training. Meta's addition of two Supreme Court litigators to its defense team suggests that the company takes the legal stakes seriously. As the case continues to unfold, it remains to be seen how the court will rule on Meta's use of copyrighted works and what impact this will have on the AI industry as a whole.

Meta did not immediately respond to a request for comment on the matter.