Destructive book scanning for AI training: where the rights of a copy's owner end

Lawfully buying a used book gives lawful access to the work, which is required for the text and data mining exception. It does not, however, address the duty to delete copies after training or create any entitlement of the author to remuneration — and this is where it is decided whether destructive scanning of books for AI training is acceptable or not.

When Markus Brandis, chair of the Association of German Antiquarian Booksellers, describes orders that within a few months earned individual second-hand bookshops hundreds of thousands of euros, he speaks of a barbaric approach. Customers from North America placed orders at night, deliberately buying up "unsellable" titles that had sat on shelves for years, and the books then went into scanners. According to an internal planning document of Anthropic that became public in a US dispute with authors, the whole acquisition was called Project Panama and its description is unusually blunt: the "intent to destructively scan all the books in the world". Spines cut off, loose pages fed into a sheet-feed scanner, the rest to recycling.

The emotions this provokes in Germany are understandable. The legal core of the matter, however, lies elsewhere than in the destruction of paper itself. Most titles processed in this way exist in thousands of copies; destroying one has no real impact on cultural heritage. The real question is whether the content of a book may be poured into a statistical model just because someone bought a used copy in a second-hand bookshop.

Lawful purchase as a ticket — but only halfway

German lawyer Christian Solmecke sums it up soberly in an analysis for DPA: scanning copyright-protected books is generally prohibited in Germany unless the author consents or a statutory exception applies. An exception exists for AI training; it is based on the EU rules on text and data mining, and its key condition is lawful access to the work. According to Solmecke, lawfully buying a copy — even a used one — is enough for this. That is also why AI companies go to second-hand bookshops rather than downloading content from pirate sources.

What is missing from this construction is interesting: remuneration for the author. The publisher and author received royalties for the original sale long ago, perhaps twenty years ago. A used copy passes through a second-hand bookshop, which collects the purchase price, and the author gets nothing from the new — commercially extraordinarily valuable — use. This is not an error of interpretation; it is an explicit feature of the rule. And it is the main reason why European legal scholarship has criticised the TDM exception from the outset as unbalanced.

Lawful purchase is therefore a ticket only to the scanning and training data preparation phase. It says nothing about what may happen to the data afterwards.

The hidden lever: the duty to delete copies

Another provision of the German Copyright Act requires copies made for training to be deleted once they are no longer needed for training. In practice, this is the most pressing point of the whole construction. Training corpora are not usually deleted after a single use — they are kept for further generations of models, for fine-tuning and for comparative benchmarks. Solmecke calls this the point where the situation becomes delicate. That is a diplomatic way of saying that a large part of the industry is probably systematically breaching this duty.

For an author or publisher, this is an essential lever that the debate so far has underestimated. The question is not only whether the company had the right to scan the book. It is whether it may still hold it in the dataset three years later, when training version 5, 6 or 7 of the model. Under the text of the German law it may not — and the evidential situation is far from comfortable for AI companies.

An interesting paradox emerged in the US proceedings against Anthropic. There, the court took the physical destruction of the book as an argument in favour of the AI company: it was said to be merely a change of format, comparable to transferring the content of a purchased CD to an MP3 player, not the creation of an additional copy. From the perspective of continental law, this reasoning is surprising — here a court would rather ask whether there was a reproduction within the limits of the exception at all and whether the other conditions were met (deletion, transparency, respect for the opt-out). Even so, this American logic shows why AI companies do not see destruction as a problem: in their reading, it is on the contrary a legal safeguard against accusations of making "additional" copies.

What this means for Czech publishers, authors and technology companies

The applicable law is determined by the place of scanning. If the books are scanned in the US — which is so far the prevailing model — local rules and US courts apply. These courts are so far divided on fair use for AI training; of three first-instance decisions to date, two came out in favour of AI companies and one against. The legal situation is open, and even a European author will have to assert their rights on someone else's turf. This is a key finding for any content protection strategy: Czech law will protect the author if the scanning takes place in the EU, but control is lost over what someone bought online and took to the US.

For publishers, this means above all working consistently with the reservation under Article 4(3) of the DSM Directive, i.e. a machine-readable opt-out from the TDM exception. Without it, the exception for commercial training applies automatically if the entity has lawful access to the work. The opt-out belongs in e-book licence terms, in metadata and in robots.txt on their own websites. If it is not there, protection is missing even before US courts come into play.

For authors, a practical guide is to review older publishing contracts. Anyone who in 2015 assigned to a publisher "all rights of reproduction and distribution in any form" now faces the question of whether this also assigned the right to license the work to an AI company. The answer is not obvious and in specific contracts often depends on how the purpose was defined. New contracts should address AI training expressly — either exclude it or make it subject to additional remuneration.

For technology companies developing their own models in the Czech Republic or the EU, the riskiest combination is unclear data provenance and no record of deletion. The AI Act requires providers of general-purpose models to publish a sufficiently detailed summary of training data and to respect authors' reservations. The argument that "the data come from purchased books", without documenting specific acquisitions and without a corpus deletion policy, will fare worse with the regulator than part of the industry admits.

Destructive scanning of books from second-hand bookshops is not the main problem of copyright in the AI era. It is just a visible symptom of the fact that the European TDM exception was designed at a time when the legislature had academic research in mind, not commercial scaling to trillions of parameters. Until the rules are revised, authors and publishers will have to protect their rights themselves — with an opt-out, precise contracts and readiness to litigate where the scanning actually takes place.

Source: Advokátní deník, 31 August 2026, "Accusations of destroying books to train AI are multiplying. What about copyright?" (in Czech; Ivana Sýkorová, citing an analysis by DPA and ČTK).