Why AI Companies Are Buying Books: The Race to Control Written Knowledge

A book can spend years on a shelf before becoming part of an AI system. In recent years, books have become valuable sources of training data for companies developing large language models. Some have acquired large collections, scanned printed pages, and converted them into digital text that can help train AI systems.

The process raises questions that go beyond technology. When books are turned into training data, who controls that knowledge, what rights do authors retain, and what happens to the physical copies after scanning? Behind the race to build smarter AI systems lies a growing debate over copyright, access to knowledge, and the future of books.

Why Printed Books Suddenly Matter to AI Companies?

AI language models learn from text, but the internet is increasingly filled with machine-generated content. A 2024 study published in Nature found that repeatedly training models on AI-generated data can lead to model collapse, a process in which models lose information and produce less reliable outputs over successive generations.

However, the risk should not be confused with a proven decline in major AI chatbots. A Techopedia analysis of model collapse notes that there is no credible public evidence that leading chatbots have degraded because of AI-generated content in their training data. Research also suggests that retaining human-generated data alongside synthetic material may help reduce the risk. The concern is therefore supported by research, but its effects on commercial AI systems remain uncertain.

This has made older printed books an attractive source of training material. Books published before the widespread use of large language models offer text created before the current surge in AI-generated content. As discussed in Eureka Street’s report on Anthropic’s book-scanning process, a book-data company has argued that pre-AI-era print books are free from contamination by AI-generated text. The appeal of printed books, then, is not simply that they contain useful information. They can also provide a source of human-written material that predates the rapid expansion of generative AI.

How Court Filings Exposed Anthropic’s Book-Scanning Project?

The practice came to light through the copyright case Bartz v. Anthropic. Unsealed court filings described an initiative known as Project Panama, under which Anthropic purchased large quantities of physical books, cut apart their bindings, and scanned their pages at high speed through commercial vendors.

According to Storyboard18’s report on Project Panama, vendor proposals covered between 500,000 and two million books over approximately six months, with projected spending reaching tens of millions of dollars. The process allowed printed books to be converted into digital text for AI training, while leaving the original copies unusable as books.

The reporting also points to a wider market for supplying books to AI companies. A metadata company has reportedly helped arrange bulk purchases of printed books under non-disclosure agreements, although publicly available details about these arrangements remain limited. 

Eureka Street’s investigation provides further context on the practice. These reports reveal how physical books can become part of the infrastructure behind AI development. They also raise questions about how copyrighted works are acquired, digitized, and used to train commercial systems.

Why Do AI Companies Destroy Books During Digitization?

Destroying the books was a choice, not a technical necessity. Google Books, for example, has used non-destructive scanning methods to digitize books borrowed from libraries, allowing the physical copies to be returned afterward. By contrast, removing a book’s spine allows its pages to be fed through high-speed scanners more efficiently, reducing the time and cost involved in processing large collections.

According to Bespacific’s report and Storyboard18’s investigation into Project Panama, court records also indicate that purchasing and scanning physical books was viewed as an alternative to obtaining large collections of pirated digital books. It also offered a way to acquire and digitize large quantities of text without negotiating individual licences for every work.

The approach highlights a practical trade-off: dismantling books can make digitization faster and cheaper, but it also destroys the original physical copies. The legal question of whether buying a book gives a company the right to scan and use its contents for AI training remains separate from the technical efficiency of the process.

The Legal Difference Between Purchased and Pirated Books

The legal dispute has two distinct parts: scanning books that were legally purchased and using books obtained from pirated digital libraries. In the US case against Anthropic, the judge ruled that converting lawfully purchased books into digital copies for the company’s own use was fair use. The ruling also found that training AI models on lawfully acquired books could qualify as transformative use under copyright law.

The treatment of pirated books was different. Anthropic agreed to a $1.5 billion settlement covering nearly 500,000 books obtained from pirate libraries. As reported by Digital Trends and TechSpot, the settlement addressed claims involving those pirated copies, rather than the destruction of books the company had purchased legally.

This distinction matters because the ruling did not establish that every form of book scanning or AI training is automatically lawful. It addressed specific circumstances under US copyright law, and the destruction of legally purchased books was a separate issue. Other countries may apply different copyright rules.

Authors’ Rights and the Control of Written Knowledge

The debate extends beyond copyright to the question of who controls access to written knowledge. After scanning, the digital copies remained in Anthropic’s private collection rather than being made available to libraries or the public. The books had been purchased, but their contents were not separately licensed from authors or publishers.

Supporters of the practice argue that purchasing a physical book gives its owner certain rights to use it and that digitization for AI training can fall within existing copyright precedents. Critics, however, raise concerns about the concentration of written material in private collections controlled by a small number of technology companies. They argue that authors may have limited control over how their work is used and may receive no direct compensation for its contribution to AI development.

The debate remains unresolved. As reported by Digital Trends and TechSpot, some authors have pursued separate legal claims rather than relying on the settlement. These disputes reflect a broader question: how should copyright law balance technological development with the rights and interests of the people who create written work?

In essence, as AI-generated content spreads, human-written books remain an important source of original text for training language models. The case involving Anthropic shows how companies can turn physical books into private digital collections, while separate legal proceedings have addressed the use of pirated material. The broader debate is not only about how AI systems are trained, but also about who controls the knowledge they learn from. Questions about authors’ rights, access to digital collections, and the destruction of physical books remain part of an evolving legal and ethical discussion.

Image Source