When we hear that companies developing artificial intelligence buy old books, remove their bindings, scan them and dispose of the paper, it can easily bring to mind a modern form of book burning.
But that comparison is not accurate.
The purpose of book burning is to suppress the content and prevent people from accessing ideas. In so-called destructive scanning, the physical medium is sacrificed precisely so that its content can be preserved digitally and processed further.
This is therefore not a story about destroying culture. It is about how human creativity becomes a raw material for artificial intelligence, and under what conditions that is lawful and fair.
What actually happened?
The best-documented case involves Anthropic, the company behind Claude.
In early 2024, it launched an internal initiative called Project Panama and spent tens of millions of dollars buying millions of used printed books. It bought them in large quantities through retailers including Better World Books and World of Books.
External contractors handled the physical processing. They removed the bindings, cut off the spines, scanned the loose pages using high-speed equipment and produced page images and machine-readable text. The paper originals were then discarded or recycled.
Why do the books have to be destroyed?
More precisely, they do not have to be. But when processing millions of copies, this is the fastest and cheapest approach.
A machine that preserves the binding must carefully turn and capture each page. Loose sheets can be fed into an automatic scanner, much like documents in a high-volume photocopier.
Rebinding a book would also often cost more than the used copy itself.
Destructive scanning is therefore not an attack on books. It is an industrial optimisation of time and costs.
Why is the internet not enough for AI companies?
The web is vast, up to date and diverse. But it is also full of duplicates, advertising, SEO content, short posts, inaccuracies, anonymous material and fragments taken out of context.
Moreover, a growing share of online content is generated by AI itself. If new models learn from the outputs of earlier models without proper controls, diversity can be lost and quality can gradually deteriorate. This phenomenon is known as model collapse.
Books offer several advantages:
- for titles published before the widespread adoption of generative AI, human authorship can be assumed with a very high degree of confidence;
- they have generally undergone writing, editing and language review;
- they contain sustained ideas, arguments and context developed over tens or hundreds of pages.
For a language model, books are therefore more than a store of words. They provide structured language and examples of how people explain, argue, tell stories and connect knowledge.
Without books and other carefully edited long-form texts, it would be much harder to build models with strong language and reasoning capabilities.
Are AI companies destroying cultural heritage?
Such a broad claim is not currently supported by evidence.
Large-scale destructive scanning organised by Anthropic has been confirmed. There have also been offers from intermediaries claiming they could supply large quantities of printed books for AI projects.
For example, ISBNdb publicly advertised such a service. After media coverage, however, it withdrew the offer and said it had only been testing market interest and had not actually bought or scanned any books for AI companies.
This demonstrates interest in such a service, not proof of another project on Anthropic’s scale.
There is no convincing evidence that all major developers use this approach. Nor is there evidence that the aim is to seek out and destroy the last surviving copies or historically rare books.
Anthropic says it obtains books from ordinary commercial sources and that its programmes do not buy or destroy rare or antiquarian books. This is, however, a company statement, not an independent audit of the entire supply chain.
With bulk purchases, mistakes cannot be ruled out entirely. If a rare or hard-to-replace copy is mixed in with ordinary books, it could be irreversibly destroyed. For now, however, there is no reason to describe that as the intended purpose of the practice.
A second-hand bookshop is not an archive either. Many mass-produced books that remain unsold for a long time would eventually be recycled or disposed of when warehouses are cleared.
Destroying one ordinary copy among thousands therefore cannot be equated with burning books to erase their content.
Where does the real ethical problem lie?
Not always in the paper itself.
A problem can arise if a rare, annotated or hard-to-replace copy is destroyed. Or if only the text is preserved, while valuable illustrations, typography, the previous owner’s notes or the significance of a particular edition are lost.
Another problem is that the digital version may remain locked in a private database. Destroying a physical copy does not automatically create a public digital archive accessible to libraries, researchers or readers.
The most fundamental question, however, is economic.
A company can buy a used book for a few euros. The author and publisher generally receive nothing from that sale. Yet the book’s content may subsequently be used to develop a product worth billions.
Such use may be lawful under certain conditions. That does not automatically make it ethical or fair to the author.
Owning the paper does not mean owning the copyright
When we buy a book, we acquire a particular physical copy. We can read it, sell it, give it away or destroy it.
We do not acquire the copyright in its text.
European law distinguishes between distributing a particular physical copy and making new copies of a work. Under certain conditions, selling a book exhausts the right to control the resale of that copy. It does not exhaust the author’s exclusive right to decide on the reproduction of the work.
This distinction also matters in the Anthropic case. The US court did not say that digitisation was automatically permitted simply because the company had bought the books. Creating a digital copy had to be assessed separately under the US doctrine of fair use.
What does “fair use” mean in the United States?
Fair use is a flexible exception under US copyright law. A court assesses the purpose of the use, the nature of the original work, the amount copied and the effect on the market on a case-by-case basis.
On 23 June 2025, Judge William Alsup ruled that using books to train models constituted transformative fair use in this particular case.
He also considered converting a lawfully purchased printed copy into a single internal digital replacement to be fair use, provided the company discarded the paper copy and did not redistribute the digital one.
This was not a general authorisation to produce and sell e-books. It was an individual judicial assessment of an internal format change and model training.
The court took a very different view of building a permanent central library from millions of e-books obtained from pirated sources such as Library Genesis or Pirate Library Mirror.
An important detail: the court did not say that every pirated book had been used for training. The unlawful acquisition and retention of those books in an internal library for possible future purposes was itself a problem.
The court rejected the argument that later transformative use automatically justified the original piracy. The case therefore headed towards a trial on liability and damages.
What exactly does the $1.5 billion mean?
Anthropic was not found liable in a final judgment, and the court did not impose a fine.
The company chose to avoid a further costly and uncertain trial and reached a class-action settlement with the plaintiffs. The court gave it final approval on 20 July 2026.
Anthropic committed to paying $1.5 billion plus interest into a non-reversionary fund. This means unused money will not return to the company but will be distributed according to rules approved by the court.
The settlement covers 482,460 eligible works. A rough calculation gives approximately $3,000 per work.
This is not, however, a sum that every author automatically receives. Approved legal and administrative costs are deducted from the fund, and payments may be divided between an author, a publisher or several rights holders.
The court approved legal fees of approximately $101.6 million, just under 6.8% of the fund, alongside further administration and distribution costs.
Anthropic also committed to destroying the original files obtained from the pirated libraries concerned and copies derived from them, except for temporary retention needed for legal proceedings.
What is the position in the European Union?
Europe follows a different legal approach.
The EU has no general fair use doctrine. Using a protected work without the author’s permission must be covered by a specific statutory exception.
Economic rights generally last for the author’s lifetime and 70 years after their death. A book published a hundred years ago may therefore still be protected if its author died much later.
Two European legal instruments are particularly important for AI training, and they address two different questions.
The Copyright Directive: when can works be analysed?
Directive (EU) 2019/790 introduced exceptions that, under specified conditions, allow works to be reproduced and extracted for text and data mining: the automated analysis of text and data in digital form.
Article 3 allows research organisations and cultural heritage institutions to carry out text and data mining for scientific research. They too must have lawful access to the works.
Article 4 creates a broader exception that may also be relevant to commercial AI training. It applies only where access to the work is lawful and the author or other rights holder has not expressly reserved that use in an appropriate manner. For online content, such a reservation is made in particular through machine-readable means.
Copies under Article 4 may be retained only for as long as necessary for text and data mining. A pirated library does not meet the lawful-access requirement.
It is also not yet definitively settled whether every stage of modern generative-model training automatically falls within this exception. European courts will gradually assess matters such as the extent of copy retention, memorisation of works and reproduction of parts of them in outputs.
The AI Act: what must a model provider demonstrate?
The AI Act does not give companies a licence to use books, and it does not create a new copyright exception.
Article 53 does, however, require providers of general-purpose AI models to put in place a policy for complying with EU copyright law. In particular, they must identify and respect rights holders’ reservations and publish a sufficiently detailed summary of the content used to train the model.
These obligations have applied to new general-purpose AI models since 2 August 2025. Models placed on the European market before that date must comply by 2 August 2027.
Article 101 has also applied since 2 August 2026. For intentional or negligent infringements, the European Commission may fine a general-purpose AI model provider up to €15 million or 3% of its total worldwide annual turnover, whichever is higher.
Legality and morality are not the same
I do not consider the destruction of one ordinary, widely available copy to be a fundamental moral problem in itself.
What matters is something else:
- whether the content was obtained lawfully;
- whether the author’s rights were respected;
- whether an irreplaceable copy was destroyed;
- whether the company can transparently explain the sources its model draws on.
Books and other high-quality human writing are extremely important to the development of artificial intelligence. Without them, building high-quality language models would be substantially harder.
But technological usefulness cannot justify everything.
I believe that in the European Union we should not prevent the use of high-quality texts in developing AI. We must, however, ensure lawful data sourcing, meaningful respect for authors’ reservations, appropriate transparency and the gradual development of workable licensing and remuneration mechanisms.
Rare physical copies whose value goes beyond the text itself deserve particular protection.
To put it in one sentence:
AI companies do not destroy books to erase their ideas. They destroy an inexpensive physical medium to obtain exceptionally valuable human content as efficiently as possible. The real issue is therefore not the paper, but the rules governing how human creativity becomes an input for commercial models.
The question today is not whether a company may destroy paper that it has bought.
It is this:
Under what conditions may a private company turn human creativity into one of the basic raw materials of its technological and business model?
#ArtificialIntelligence #Copyright #Books #AIAct