Out of Concern for Copyright? AI Companies and Their Destructive Scanning of Books
Books are a valuable part of our cultural heritage. They are preserved and made accessible in libraries (in this case, the SUB Göttingen). Foto (ed.): mibuchat, CC BY-SA
The Destructive Scanning of Books
The retro-digitization (usually referred to simply as “digitization”) of books is typically carried out using scanning technology. Modern book scanners are capable of scanning books non-destructively. However, non-destructive scanning of books is a complex process, as irregularities caused by the book’s binding must be compensated for during the scanning process.
In contrast, in so-called destructive scanning, the book’s binding is mechanically removed. This turns a book into a loose collection of pages that can be conveniently digitized using a sheet-fed scanner. Once the scanning process is complete, the book pages are then disposed of as waste paper.
Why Antiquarian Bookstores Are Important for AI Training
Digital data sources, such as web archives, have long been incorporated into the training of current AI models. The digitization of printed books is now the next stage in data acquisition. According to observers, tech companies are particularly focused on non-fiction books published between the 1970s and the 1990s.
These books are likely to be of interest for three reasons: First, nonfiction books are valuable for AI training because of their content; they contain factual information that is still largely accurate and is integrated into linguistically sophisticated patterns of argumentation. Second, book titles published before the mid-1990s were generally released exclusively as print editions and have not been digitized since. Third, books from the period in question are still predominantly subject to copyright, which expires only 70 years after the author’s death in both Germany and the U.S.
Obtaining scans of these books from dubious third-party sources would be illegal. Therefore, from a copyright perspective, the purchased print editions of these books are particularly valuable to the companies.
Technical Reasons for Destructive Scanning
But why do AI companies accept the destruction of the books during the scanning process?
First, technical aspects play a significant role in large-scale retro-digitization of books: As explained at the outset, destructive scanning is simpler, faster, and therefore more cost-effective. Although the AI industry appears economically strong, observers interpret the decision to use the most cost-effective digitization method as an indication of cost pressure in this “highly competitive industry.”
Competitive considerations may also play a role, because while the books in question are not particularly old or unique, many of them are nevertheless rarely available on the antiquarian book market due to low print runs and scarcity. Whoever buys them first effectively removes them from the competition’s reach, thereby gaining a small advantage for their AI model.
Destructive Scanning Out of Concern for Copyright?
A lawsuit between three plaintiff authors and the AI company Anthropic before the United States District Court for the Northern District of California (Bartz v. Anthropic) established a legal precedent regarding destructive book scanning in 2025.
The central question here was whether destructive scanning, on the one hand, and the training of AI models using the content of acquired digitized books, on the other, fall under the so-called “fair use principle.”
The fair use principle is a legal doctrine found in some common law countries. It permits the use of copyrighted material for purposes of public education or intellectual production without requiring authorization for such use. Classic areas of application for fair use include commentary, criticism, parody, news reporting, research, and scholarship. Under German copyright law, instead of a general clause like fair use, there are narrow exceptions that permit certain uses (such as quotations, caricatures, parodies, educational and scholarly purposes, reporting on current events, etc.) regardless of the copyright holder’s consent.
The two central questions to be decided by the court must be considered separately:
First, whether training AI models with the content of purchased digitized books is protected by the fair use doctrine.
Second, whether destructive scanning constitutes a copyright-relevant copy of a work.
According to Judge William Alsup, training AI models using the content of purchased digitized books was “transformative in the truest sense of the word” and therefore protected by the fair use doctrine. Destructive scanning should be classified as “replacement scanning,” in which a printed work is replaced by a digital work. This conversion of a purchased printed book into a digital book was classified by the court as a “format change” rather than a copy and is therefore also covered by fair use.
A Landmark Ruling with Significant Implications
The rule on format conversion—also known as the “format-shifting rationale”—thus promises, according to this landmark ruling, a certain degree of legal certainty for AI companies that destroy the books they have acquired for AI training through destructive scanning. The method that is more efficient from a technical standpoint is therefore also the more legally advisable one.
However, it is a fallacy to assume that the court’s ruling requires AI companies to destroy the digitized books. The lawsuit merely examined the specific digitization process as applied by Anthropic as part of “Project Panama”—and classified it as fair use.
E-books Instead of Print Books?
If e-books were used for training instead of their printed counterparts, the time-consuming step of scanning—and the associated issues of transportation, storage, and reuse—would be eliminated.
However, e-books can only become training data for AI models in two ways: the legal way—that is, with the involvement of the rights holders of those e-books—or by training with pirated copies. For AI companies, the legal route entails a significantly higher “legal/practice/business slog” (quote from Anthropic co-founder and CEO Dario Amodei).
Training AI with pirated copies of e-books is not covered by fair use and thus constitutes a clear copyright infringement, Judge Alsup ruled.
Fears of ‘Fahrenheit 451’ Scenarios
Reports of AI companies purchasing large quantities of books from used bookstores have sparked, in some cases, fierce reactions in the media. The media discussion often references Ray Bradbury’s novel *Fahrenheit 451*, a dystopia in which books are banned.
A closer look at the relevant court ruling in this specific case reveals that there is no legal obligation to destroy retro-digitized books. Economic reasons for choosing the most cost-effective digitization method therefore appear to be the decisive factor when the AI industry employs destructive book scanning. Against the backdrop of the actions taken by major AI companies, calls for open-source AI models are currently growing louder, and with regard to the retro-digitization of books, there are also increasing calls for “an open-source initiative or a similar nonprofit project.”
Would you like to support iRights.info?
iRights.info provides information and explanations on the subject of “Copyright and creativity in the digital world”. All texts are published free of charge and openly licensed.
If you like, you can support us via the donation platform Betterplace and receive a donation receip. Betterplace accepts PayPal, direct debit, credit card, paydirekt or bank transfer.
We would be particularly pleased to receive a regular contribution, for example as a monthly standing order. The non-profit organization iRights e.V. would like to thank you for your support!
DOI for this text: https://doi.org/10.59350/ksfmn-23248 · Automatic DOI assignment for blogs via The Rogue Scholar
Was sagen Sie dazu?