A technical book no one has checked out in years. A monograph published in a small print run in the ’80s. A medical textbook that’s long been out of circulation. Until recently, such books were mostly a problem for libraries and antiquarian bookstores: they took up space, it was hard to find a buyer, and digitization often just wasn’t worth it.<\/p>
The development of artificial intelligence is changing this landscape.<\/p>
In Poland, the issue has surfaced in recent days after reports about foreign companies interested in buying books from antiquarian bookstores. PAP described cases of such offers<\/a><\/mark>, linking the interest in books to the growing demand for data among AI model developers.<\/p> We also don\u2019t know exactly what happens to every book bought in Poland today or where it ultimately ends up. So there\u2019s no basis for claiming that Polish collections are being shipped en masse to AI labs. We do know, though, that the mechanism itself is perfectly plausible. One of the largest companies developing generative AI has already done something very similar on an industrial scale.<\/p> We’re talking about Anthropic, the producer of the Claude models.<\/p> In the U.S. case Bartz v. Anthropic<\/em>, book authors accused the company of copyright infringement during the creation of datasets used to develop AI models. The federal court for the Northern District of California provided the public case records<\/a>. The documents indicate that Anthropic built its own digital library in two ways. The company used books previously obtained from unauthorized online libraries, while at the same time it began buying vast quantities of print copies. The court later treated these two practices differently.<\/p> The process for legally purchased books was surprisingly analog for one of the world’s most advanced tech businesses. Anthropic spent millions of dollars on millions of copies, including used books. Then external contractors removed the bindings, cut the books apart, scanned the pages, and created digital copies of the text. After that process, the paper copies didn’t go back on the market.<\/p> Some of the library built this way was later used to create training datasets for models. <\/p> At first glance, the whole undertaking seems rather absurd. Tech companies have had the internet, digital libraries and massive text corpora for years. Why buy truckloads of paper books, cut them up, and scan them?<\/p> One obvious answer is this: a large share of human knowledge has never made it onto the internet in a form you can just download and use. This is especially true of older, specialized, local publications and those published in languages other than English. Old textbooks, industry reports, academic dissertations or regional monographs may exist in libraries and private collections while being virtually absent from the open internet.<\/p> But there\u2019s another problem. The internet itself is changing.<\/p> The first large language models emerged at a very particular moment. Back then, the vast majority of text online was undeniably human-written. Blogs, forums, Wikipedia, news sites, documents, comments and websites were created before generative AI started producing text at scale.<\/p> These days, you can’t make that assumption anymore.<\/p> Models generate articles, product descriptions, marketing copy, comments, SEO posts and entire websites. So another model that collects data from the web is increasingly running into content that other models created earlier.<\/p> That doesn’t mean synthetic data are bad by definition. The problem comes when generated data start replacing real-world data instead of complementing them.<\/p>Millions of books on the chopping block<\/h4>
The internet isn’t enough<\/h4>