Anthropic, Amazon, and Meta have been purchasing and destroying physical books to train their large language models (LLMs), sparking legal and market disruption concerns. This practice, while legal, has raised questions about copyright and the future of the book market.
TL;DR
- Anthropic, Amazon, and Meta are buying and destroying physical books to train their AI models, aiming for high-quality human-generated text.
- The practice is legal under the first sale doctrine, but raises concerns about copyright and market disruption.
- Booksellers have noticed unusual orders, with some suspecting AI companies behind the bulk purchases.
What happened
Anthropic, Amazon, and Meta have been buying large quantities of physical books to train their AI models. These books are scanned and then destroyed to avoid legal issues related to distributing digital copies.
Booksellers across the country, including in Mississippi, have reported unusual orders that they suspect are related to AI training. Columbus-based bookseller David Hallinan received several orders in May that he believes were part of this AI buying spree.
The orders were placed through AbeBooks, a massive e-commerce book marketplace owned by Amazon. Many sellers suspect these orders are part of Anthropic's 'Project Panama', which involved bulk-buying books, scanning them, and then destroying the originals.
Why it matters
This practice raises concerns about copyright and market disruption. While destroying a purchased book is legal under the first sale doctrine, the potential distribution of copyrighted works could threaten the market.
AI companies are turning to physical books for data quality and legal liability. Since 2022, large amounts of modern text contain AI-generated content, which cannot be used to train the models further.
The lack of regulation in the AI industry adds to the uncertainty. Law professors have noted that the law is not equipped to handle the new developments in AI, and the speed and scope of LLMs consuming materials could have unforeseen consequences.
Key facts
- Anthropic, Amazon, and Meta are buying and destroying physical books to train their AI models.
- Booksellers have reported unusual orders, with some suspecting AI companies behind the bulk purchases.
- The orders were placed through AbeBooks, a massive e-commerce book marketplace owned by Amazon.
- Many sellers suspect these orders are part of Anthropic's 'Project Panama', which involved bulk-buying books, scanning them, and then destroying the originals.
- The practice is legal under the first sale doctrine, but raises concerns about copyright and market disruption.
- AI companies are turning to physical books for data quality and legal liability.
- Law professors have noted that the law is not equipped to handle the new developments in AI.
- The speed and scope of LLMs consuming materials could have unforeseen consequences.
Context
The use of physical books to train AI models is a relatively new practice, driven by the need for high-quality, human-generated text. As AI models become more advanced, the need for quality training data becomes increasingly important.
The legal landscape surrounding AI and copyright is still evolving. While the first sale doctrine allows for the destruction of purchased books, the potential distribution of copyrighted works remains a concern.
The lack of regulation in the AI industry adds to the uncertainty. As AI models continue to consume and redistribute materials, the legal and market implications will need to be carefully considered.
