**Editor’s Note**: This analysis forms part of The Atlantic’s look into the Library Genesis dataset. Access The Atlantic’s search tool for movie and television writing used to train AI here. When Meta began creating its AI model, Llama 3, the team confronted an ethical dilemma: acquiring a vast amount of quality text legally was slow and costly. They considered pirating data after being dissatisfied with licensing options, citing high costs and delayed delivery from potential partners. Internal discussions revealed a strong desire to use books, viewed as essential for training, prompting them to explore Library Genesis (LibGen), a large repository of pirated literature and research. Internal communications revealed Meta employees sought approval from CEO Mark Zuckerberg to utilize LibGen’s dataset, now public due to a copyright lawsuit from several authors. Moreover, OpenAI has been associated with LibGen in past use cases as well. Although the full scope of texts both companies trained on is unclear—LibGen's content constantly updates—the database contains millions of titles, including significant works and academic papers. Meta and OpenAI defend their approaches under the assertion of “fair use, ” arguing that their generative AI transforms sources into new content. However, LibGen's utilization raises critical issues, particularly since internal documents indicate Meta downloaded data via BitTorrent, which carries legal risks due to potential distribution of pirated content.

Meta has insisted it took precautions against seeding files. Employees acknowledged the legal risks, discussing strategies to hide their activities, such as avoiding references to copyrighted works and filtering out identifiable information. LibGen, significantly larger than other pirate collections, appeals to AI developers due to its broad selection, including contemporary literature and academic journals. Established in 2008 by Russian scientists, LibGen serves areas with limited access to educational resources. Its growth has been fueled by contributions of pirated materials, leading to a predominance of English texts over time. Despite multiple attempts by publishers to curb piracy, including significant court rulings and fines against LibGen, the repository persists. This accessibility raises ethical concerns regarding the underlying work of authors, who often do not receive credit or compensation. Generative-AI technologies risk decontextualizing knowledge and undermining the recognition due to original creators. The central challenge remains how to effectively balance the dissemination of knowledge and creative work for societal benefit, as companies like Meta capitalize on these resources for profit, potentially diminishing the value of human intellectual engagement.