The Artificial Intelligence (AI) race has a new victim — second-hand books. Over the past summer, reports have emerged of AI developers around the world purchasing vast quantities of used books. Once acquired, these volumes are destructively scanned: Their bindings are cut off and the resulting loose pages are fed through automated scanners, rendering the physical volumes unusable.
![Much of India’s intellectual life between 1800 and 1950 was conducted through the medium of periodicals. Because of poor management and budgetary constraints, libraries in India allowed their collections of these periodicals to deteriorate. [Photo credit: Mint Archive]](https://www.hindustantimes.com/ht-img/img/2026/09/04/550x309/Lead_Photo_1788541790632_45db0560-14f7-4d2c-b4d4-cd2ecbeda1e3_1788541796439_ea49e3e2-8042-4893-9117-5811290225dc_0f794d34-3c89-4e16-a7da-342e3f3c6a9a_9c7813c8-2280-475f-b638-09bc7c4cb44f.jpg)
This interest in used books stems from two factors. The first is technical. A growing body of literature on model collapse suggests that recursively training AI models on slop (or AI-generated data) can progressively erode their ability to track distributions. The problem is commonly compared to making photocopies of photocopies, a process that produces ever blurrier results. This insight has led AI firms to prize authentic human-authored text. But, as it is becoming harder to discern whether the text we encounter today has already been shaped by AI, developers have become interested in tapping an uncontaminated source — books published before the large language model (LLM) era.
ALSO READ | India's GDP: When no number high enough is enough
The second factor is legal. In the landmark case Bartz v. Anthropic (2025), Judge William Alsup ruled that Anthropic violated the law when it downloaded pirated copies of books from websites like LibGen. Anthropic was, however, entitled to destructively scan books it had lawfully acquired because it was effectively exchanging one format (print) for another (digital). It was also “fair use”, Judge Alsup ruled, to utilise the resulting scans for “training”, since “generative” AI is “quintessentially transformative” — it seeks not to “replicate” works but to generate “new text”. This decision encouraged AI firms, especially those operating in the US, to begin acquiring physical copies of the millions of books they need to train their models.
{{/usCountry}}The second factor is legal. In the landmark case Bartz v. Anthropic (2025), Judge William Alsup ruled that Anthropic violated the law when it downloaded pirated copies of books from websites like LibGen. Anthropic was, however, entitled to destructively scan books it had lawfully acquired because it was effectively exchanging one format (print) for another (digital). It was also “fair use”, Judge Alsup ruled, to utilise the resulting scans for “training”, since “generative” AI is “quintessentially transformative” — it seeks not to “replicate” works but to generate “new text”. This decision encouraged AI firms, especially those operating in the US, to begin acquiring physical copies of the millions of books they need to train their models.
{{/usCountry}}But why scan books destructively? There is a logic to the brutality. Unbound volumes can be scanned significantly faster. Because loose pages can be scanned flat, they are also far less susceptible to distortions such as curves and shadows, improving image consistency and optical character recognition. Altogether, this means that destructive scanning is notably cheaper than non-destructive methods (by some estimates, it costs 50-75% less per page). This combination of higher speed and accuracy at lower cost is appealing because AI firms are in an arms race, and the better a model is trained, the better its output will be. From the spoils comes the victor.
What should we make of this development? There is no cause for concern when it comes to books that are in plentiful supply. But it becomes a different matter when AI firms seek to acquire and destructively scan works that are rare — as a recent investigation by the technology publication 404 Media has alleged. It is not hard to see why AI firms may be tempted to acquire “uncommon books”. Since the performance of AI models depends on the breadth and depth of their training, it is likely that, as internal correspondence disclosed during discovery in Bartz v. Anthropic revealed, AI firms may eventually seek to “destructively scan all the books in the world”.
This ambition raises three concerns. First, it becomes possible that all publicly available copies of a rare or scarce book could be inadvertently destroyed. AI firms have downplayed this risk by stating that they would never destroy “collectibles”. But this ignores a coordination problem. Suppose only five copies of a particular work are known to survive. Several firms could purchase and destroy them independently, each believing that other copies remained intact. No individual purchase would appear catastrophic, yet the cumulative result could eliminate every known copy.
ALSO READ | India’s gig workers need more effective protection
Second, even if collectibles are not destroyed, they may nevertheless become unavailable to readers. For instance, Elon Musk has promised that SpaceXAI will preserve any rare books it acquires. But where? It is not desirable that collectibles end up in corporate vaults. This worry is not theoretical. Google Books contains a vast array of rare works deemed to be in the public domain in the US. But those works are not always viewable by users located outside the US. This is because Google Books worries that those rare works may not be deemed to be in the public domain in foreign countries, and so to minimise its own liability, it pre-emptively limits access for foreign users. As a result, one can end up being able to view a work published by an Indian in India in 1875 — only if one is located in the US.
Third, and perhaps most worryingly, because AI firms operate across national and cultural boundaries, they may not always sufficiently appreciate the rarity or value of the works they encounter. For instance, much of India’s intellectual life between 1800 and 1950 was conducted through the medium of periodicals. Because of poor management and budgetary constraints, libraries in India allowed their collections of these periodicals to deteriorate. Consequently, the last surviving copies of many of these periodicals can now only be found overseas, which includes the stock rooms of antiquarian dealers. A book procurement system optimised for scale may destroy such items of cultural significance without even recognising what has been lost.
It is essential, then, that AI firms exercise due care. In the first instance, they should refrain from destructively scanning works published before the era of mass printing (which, for the sake of convenience, we can assume began in 1950). In such cases, destructive scanning should proceed only when reviews based on catalogue research and consultations with archivists and scholars confirm that a genre or body of work is not at risk of extinction. Furthermore, AI firms should create a registry of target acquisitions so that competition does not inadvertently cause rare works to become extinct. Finally, seeing as they are reducing the supply of used books, AI firms should commit to making every out-of-copyright work they digitise freely available on the web. This would also reduce the need for their competitors to procure additional scarce copies of the same work.
All this goes to say that the ceaseless appetite of AI firms for training data can be made to work for the good of humankind. The digitisation they undertake can help guard against unexpected losses caused by fire, flood, war, natural disaster, theft, decay, and institutional neglect. It can also make it possible for vastly more readers to access and engage with books that are scarce or fragile or located in distant repositories. But whether this happy day comes to pass depends on the careful management of the business of scanning “all the books in the world”. The question before AI firms today is whether they will treat rare and valuable books as disposable training inputs or as the spines of civilisation.
Rahul Sagar is a Global Network professor at NYU Abu Dhabi. His latest book is The Birth of Indian Liberalism. The views expressed are personal