AI company targets physical books and exports 563 tons of books to the United States

Recently, news has spread that dozens or hundreds of metric tons of books have been acquired in Taiwan, Japan, and South Korea. The acquisition channels include publishing dealers or second-hand bookstores. It is speculated that it is highly related to the "Project Panama" of the American AI company Anthropic.

South Korea's "Central Daily News" pointed out that transportation records from November 2024 to the present show that Anthropic purchased books from Taiwan's United Book Distributors (UDC) in 16 batches, exporting a total of 563 metric tons of books. The latest batch of orders arrived in the United States on September 4. It is worth noting that the document mentions "Simplified Chinese books from China," indicating that Anthropic may indirectly import Chinese books through Taiwan.

This year, a large number of second-hand book orders began to appear around the world. Germany, Japan, South Korea, etc. have all received large orders from mysterious buyers. South Korea has received procurement inquiries since August, hoping that bookstores can regularly supply Korean books on a pallet or container scale. Japan received an unusual number of orders in July and August, totaling more than 50 metric tons and about 100,000 books, indicating that bookstores or dealers may be using works as training materials for AI companies without their knowledge.

The physical book is edited, has a complete content structure, and can supplement limited online information, which is helpful for AI-generated quality. The collected books will be cut, scanned, and then thrown into a shredder for destruction. When the scope of acquisitions extends to out-of-print books and rare publications, it raises concerns about cultural preservation in the publishing industry.

Anthropic, an American AI company, was previously involved in the Bartz case, an infringement lawsuit filed by three book authors against the unauthorized use of books to train large language models. The court ruled that if the training materials were purchased with money, scanned page by page, and created a digital file to train the AI, since the content of the work is not reproduced or leaked, it complies with fair use under U.S. copyright law. If the training data comes from a pirated database, the source itself is illegal and cannot constitute fair use.

The case was eventually settled with a compensation of US$3,000 per book. Nobel Prize winner Han Jiang and many Korean writers were also eligible for compensation. However, it should be noted that the scope of the settlement does not include potential infringement issues of AI-generated content. In the cross-border book purchasing network of AI companies, there are huge problems posed by artificial intelligence, law, and the publishing industry. In addition to the damage and disappearance of rare ancient books and out-of-print items, and concerns about cultural preservation, the writing laws for training AI may need to consider future dimensions and explore potential usage issues.