In recent weeks, booksellers in Australia and Europe have raised concerns about unusual, bulk purchases of books, amid reports that artificial intelligence companies are acquiring physical copies to train their datasets.However, it is the manner of such training that has caused alarm: unsealed documents in a copyright lawsuit against Anthropic revealed that the company was resorting to ‘destructive scanning’, a process in which books are destroyed after their contents are scanned.The Guardian reported that Australian booksellers had received requests for books ranging from rare and obscure titles to cheap paperbacks through intermediaries, while Fortune reported a similar incident involving a Dutch bookseller. Several of these books do not have known copies available, making the practice all the more egregious.In many cases, the purchases are not made directly by AI companies. Instead, intermediaries source books in bulk from booksellers before sending them for digitisation, making it difficult for sellers to know who the eventual buyer is or how the books will be used.The reports build on earlier findings about Anthropic’s Project Panama. In January, The Washington Post reported that the company had bought millions of physical books, removed their bindings and scanned them to build a massive collection of high-quality training data for its models.The new reports suggest that a practice once visible mainly through court filings and investigative reporting is now becoming visible further up the supply chain, with booksellers noticing unusual demand. We explain.Why books?Conventional use of the terms AI and generative AI refers to a subset of artificial intelligence called large language models (LLMs). These models learn statistical patterns from enormous volumes of text input and generate textual (or visual) output based on the quality of data fed into them.Story continues below this adWhile the competition in recent years has largely centred on acquiring maximum computing power and accessing the chips that drive them, the scramble for physical books reflects a growing demand for another critical resource: high-quality training data.For years, frontier AI companies have relied heavily on text scraped from the public internet. But as AI-generated content has proliferated online and copyright disputes have made large-scale web scraping increasingly contentious, companies have begun looking elsewhere for reliable, human-written material. Books offer long-form, edited text across a wide range of subjects, much of which is either unavailable or difficult to obtain in digital form.Also in Explained | Is everything on the internet now written by AI? The science of AI detection tools, how efficient they areISBNdb, which markets book-sourcing services to AI companies, has argued that books published before the generative AI boom are particularly valuable because they are free of AI-generated content. As AI-generated text proliferates online, older books are increasingly seen as a source of cleaner training data.Jannis Lennartz, Visiting Professor at Humboldt-Universität zu Berlin, argues in an essay for Verfassungsblog titled AI Is Eating the Book World that the current legal debate over book digitisation began with Google Books, Google’s effort to digitise millions of library books in the 2000s. The project sparked years of litigation over copyright and mass digitisation, becoming the defining legal battle over large-scale book scanning.Story continues below this adLennartz argues that the current digitisation efforts by AI companies resemble Google Books in technique but differ in purpose. Instead of making books searchable or accessible online, the goal is to convert them into machine-readable training data for AI systems. He says that the present demand for books also changes the economics of digitisation, as books are increasingly acquired not for reading or preservation, but for AI training.Many books are unavailable as clean digital files, while publishers often control access to commercial ebooks. Physical copies therefore remain one of the simplest ways to obtain large volumes of text for digitisation.The present controversy extends beyond the acquisition of books to what happens to them during digitisation. Once acquired, the books must be converted into machine-readable text before they can be used for training.Also Read | Can AI-driven growth be made responsible? An Expert Explains howMany AI companies have turned to destructive book scanning, a process that permanently disassembles physical books for their high-speed digitisation. In the process, they cut the spines of books and feed the loose sheets to the automated document scanners, choosing to recycle or shred the remaining pages. The scanned images are then converted into text using optical character recognition (OCR), allowing AI models to ingest them during training.Story continues below this adDestructive scanning has long been used in some mass-digitisation projects because it is faster and produces flat, high-quality images for optical character recognition. However, preservation projects often favour non-destructive scanning to avoid damaging physical books.The Internet Archive, a digital library best known for preserving web pages and digitising books, has long maintained that it uses non-destructive digitisation, photographing books without dismantling them to preserve the original copy.In an email reply to The Indian Express, Chris Freeland, Director of Library Services at Internet Archive, said the organisation’s 2021 explanation of its scanning process “continues to be the best description of its digitisation process today”.In that post, the Internet Archive said its testing of commercial book scanners, which feature a vacuum-powered page-turning arm, threatened the integrity of brittle books, rare volumes and other special collections that its library partners wanted digitised.The legal questionsStory continues below this adThe acquisition and digitisation of physical books for AI training also raises a fresh set of copyright questions, even if the books themselves were lawfully purchased.Arul George Scaria, professor at the National Law School of India University, argues that buying physical books and digitising them for AI training represents a middle ground between negotiating licences with publishers and relying on shadow libraries.“Buying physical copies and getting them digitised for training appears to be the middle path taken by many AI developers now, as many presume that this is fairer to the authors and publishers,” he told The Indian Express.The legal position under Indian copyright law remains unsettled. “Whether this is permissible under copyright law may depend on different factors including the question of whether digitisation is done solely for the purpose of AI training or whether the firms had stored it for other purposes also,” Scaria said. “If it is done solely for the purpose of training, it should fall under the fair dealing exception under the copyright law.”Story continues below this adThe fair dealing exception, outlined in Section 52 of the Copyright Act of 1957, is a copyright provision allowing the limited use of protected material without the owner’s permission for specific purposes such as private study, research, criticism, review, and news reporting.Scaria, however, argued that copyright law alone may not resolve the concerns raised by authors and publishers. “Many copyright holders have anxieties and concerns about the use of their works,” he said, adding that “we should also explore solutions outside copyright law”, including mandatory labelling of AI-generated content and funds to support authors and artists affected by large-scale AI adoption.