In July of this year, a batch of used books was secretly fitted with AirTags and shipped from California to Las Vegas, eventually ending up in an Amazon warehouse, where they were discovered to be used for destructive scanning—spines were cut off, pages were scanned, and then the books were destroyed. After media exposure, this sparked the "AI book burning" controversy, with fingers pointed at AI companies' large-scale acquisition of used books for training models.

Recently, multiple used book dealers in the UK, Ireland, and Australia have discovered that mysterious buyers are purchasing large quantities of obscure old books at any cost and without regard to content. The orders include local histories, translations in minor languages, specialized manuals, and other books that have long been ignored. A UK seller said that since 2026 alone, such buyers have ordered about 6,000 books. These books are ultimately shipped to the same warehouse, where they undergo spine-cutting and scanning.

Investigations show that the "Panama Project" of the US AI company Anthropic specializes in such "destructive scanning," investing tens of millions of dollars to purchase millions of paper books to expand training data. A court ruled its actions as "fair use," determining that legally purchasing and destroying paper books and replacing them with digital copies does not constitute infringement. However, public opinion widely fears that this move may accelerate the disappearance of out-of-print books.

AI companies' large-scale purchase of used books also reflects the demand for high-quality training data. As AI-generated content floods the internet, models are said to face the risk of "model collapse," while old books are regarded as precious corpus due to the originality of their content and the seriousness of their writing. Many creators are resisting through watermarks, API blocking, or "strictly prohibit AI training" declarations, and some publishers have also explicitly marked on copyright pages that AI use is prohibited.

To respond to this trend, the Authors Guild of America launched a "Human Authored" certification, with about 3,000 authors completing certification for 5,000 books. In July 2026, a court approved Anthropic to pay $1.5 billion in settlement to authors and publishers, but this settlement only covers training data based on pirated websites and does not include paper book scanning.

The controversy over "AI book burning" exposes the conflict between commercial interests and the traditional text ecosystem in AI data collection. The knowledge inheritance carried by used books is facing new tensions in the data-driven era.

Source: AI Companies Mass-Purchase Used Books for Destructive Scanning to Train Models, Sparking "AI Book Burning" Controversy · Originally published on 2026-08-29