AI devours books for training, sparking concerns over printed heritage

Artificial intelligence developers are turning to physical books as a new source of data to train their language models, aiming to escape the "contamination" of online text generated by AI itself. The scarcity of clean, human-written data on the internet has pushed these companies to acquire large quantities of printed books.

To accelerate the data intake process, companies employ "destructive scanning." This method involves using hydraulic guillotines to cut the spines off books, allowing pages to be fed through industrial scanners at high speeds. The paper is then shredded or recycled. This approach is driven by the urgency to obtain high-quality, human-authored text before the widespread adoption of chatbots degrades available online content.

Companies are reportedly making significant purchases, with one entity managing sales of up to a million copies under strict confidentiality agreements, according to ISBNdb. Booksellers in the Netherlands and Germany have noted unusual purchases of specialized texts, as reported by NL Times. In the United States, courts have upheld this practice, ruling in a lawsuit against Anthropic that digitization for training purposes constitutes legal "transformative use" because the original books are destroyed and the copies are not publicly distributed.

Despite legal backing and technical rationale, the mass destruction of books is causing apprehension among authors, bibliophiles, and booksellers. They view the diminishing printed heritage as becoming disposable input for the digital age, sparking debate about the value and preservation of physical literature.