AI companies may be pulping rare books to cut training costs, though non-destructive alternatives exist

The practice of AI companies purchasing and destroying books to fuel model development has raised alarm among booksellers and literary preservationists, who worry that irreplaceable texts are being lost in the race to build better systems.

The mechanics of the problem#

Tech firms pursuing high-quality training data have turned to old books as a source of coherent, engaging long-form text—the kind of material that helps train more capable language models. The fastest and cheapest way to extract this content is to physically destroy the books themselves: tearing pages, feeding them through scanners, then discarding the material. For companies operating under pressure to advance their models quickly, this represents an economical shortcut that leaves nothing behind but data.

The concern extends beyond isolated incidents. Observers suspect the scale of book destruction is far larger than currently documented, and that permanent loss of physical copies is already occurring, particularly among rare and out-of-print titles.

A working alternative already exists#

The technology to avoid this damage has been available for years. Google patented a non-destructive book-scanning method in 2009 that could allow AI firms to digitize texts without harming originals. The technique uses curved glass and careful page handling to capture content while keeping books intact.

The approach is not without limitations:

• Page curves can distort captured text
• Pages may be missed during the scanning process
• Workers must move carefully, as rushing introduces errors like hands obscuring page content

Why companies may not adopt it#

Despite its viability, the non-destructive method requires time, care, and capital investment—costs that destructive scanning eliminates entirely. For firms racing to scale their training datasets, the economic calculus favors speed and affordability over preservation.

Institutions like the Internet Archive have long recognized that handling rare and aging materials demands patience and precision. Their approach to preservation underscores that meaningful digitization is inherently labor-intensive—a reality that sits uneasily with the economics driving book acquisition for AI training.