Meta openly uses pirated books for AI training with Zuckerberg’s approval

Meta and other major AI companies are openly using pirated book collections to train their AI models, creating a growing tension between technological advancement and copyright protection. This controversial practice reveals how AI developers are prioritizing rapid development over legal considerations in the race to build more capable large language models, raising significant questions about ethical data sourcing in the AI industry.

The big picture: Meta employees received permission from CEO Mark Zuckerberg to download and use Library Genesis (LibGen), a massive pirated repository containing over 7.5 million books and 81 million research papers, to train their Llama 3 AI model.

• Internal discussions revealed that Meta employees considered licensing books and research papers but found the process “unreasonably expensive” and “incredibly slow,” taking “4+ weeks to deliver data.”
• Meta employees acknowledged that training Llama on LibGen presented a “medium-high legal risk” and discussed various strategies to mitigate or mask their activity.

Why this matters: The use of pirated materials reveals the enormous data hunger driving AI development and the ethical corners being cut to feed advanced models like Llama 3 and ChatGPT.

• AI companies argue their use of copyrighted works constitutes “fair use” because large language models (LLMs) “transform” the original material into new work.
• This practice creates a fundamental tension between rapid AI advancement and respecting intellectual property rights.

Historical context: LibGen originated around 2008, created by scientists in Russia primarily to serve people in regions with limited academic access.

• The collection has expanded dramatically over time, shifting from primarily Russian-language works to predominantly English-language content.
• It has become one of the largest repositories of pirated intellectual content in existence.

Between the lines: The willingness of major tech companies to use pirated content indicates how competitive pressures are shaping ethical decision-making in AI development.

• Despite acknowledging legal risks, companies appear to be calculating that the benefits of using this data outweigh potential legal consequences.
• The practice suggests that current copyright frameworks may be insufficient for addressing the unique challenges posed by AI training.

The bottom line: As AI development accelerates, the industry faces growing scrutiny over its data practices, forcing a reckoning with fundamental questions about knowledge ownership, fair compensation for creators, and appropriate boundaries for machine learning.