Anthropic bought millions of books, cut them apart for AI training — now US groups want a probe

US civil society groups are pushing back against a practice by AI companies of buying physical books in bulk, scanning them for model training data and then destroying the originals. The groups say Big Tech firms, armed with vast capital, are wiping out limited book supplies to shut out smaller competitors.

On Friday, 18 organizations — including Demand Progress Education Fund and the Consumer Federation of America — sent a joint letter to Federal Trade Commission Chair Andrew Ferguson and Commissioner Mark Meador urging an immediate investigation into the practice.

The groups said AI companies' bulk purchase of physical books, followed by scanning and permanent destruction of the originals, constitutes an unfair exercise of market dominance.

They said the mass destruction of books amounts to "source foreclosure" — artificially blocking later-stage startups from entering the market. The groups urged the FTC to open an investigation under federal statutes prohibiting unfair competition.

"AI companies are permanently destroying books en masse, cutting off broader public access to these resources," the groups said. "This ensures that only the wealthiest incumbent firms can build high-quality AI models."

They added that by destroying original works, these companies are "creating a future in which they alone remain as the exclusive custodians of humanity's recorded knowledge."

Anthropic has in fact been running what it calls the "Panama Project" to build AI training datasets — buying millions of books, slicing off the spines, scanning the pages and recycling the damaged copies as waste paper.

Amazon is also training its AI models in a similar fashion, according to reports.

IT publication 404 Media reported that a bookstore receiving a bulk order of 1,000 volumes attached tracking devices to the books and confirmed they were delivered to a book-scanning facility inside an Amazon warehouse in Las Vegas.

However, a court has previously ruled that using legally purchased books for AI training does not in itself violate copyright law. A US federal court, in a copyright lawsuit brought by authors against Anthropic, found that using legally purchased books for AI training qualifies as "fair use."

The civil society groups maintain that even if the practice does not breach copyright law, it may still constitute unfair competition.