AI Is Scanning the World's Books. It Could Be Preserving Them Too

Editorial image generated by the author.

There is something instinctively disturbing about cutting the spine off an old book, feeding its pages through a scanner and throwing what remains into recycling.

That reaction becomes stronger when the purpose is training an artificial intelligence model. Anthropic did this on an industrial scale under Project Panama, purchasing large numbers of physical books, removing their bindings and scanning them into a digital library. More recently, 404 Media traced a shipment containing rare books to an Amazon facility that its investigation identified as an AI-training operation, where employees described books being disbound and digitised.

It makes for an effective story about the excesses of the AI industry: immensely valuable companies consuming physical remnants of human culture to feed their machines. But the outrage obscures a more interesting problem. The problem is not that these books are being digitised. Many of them probably should be. It is that companies can spend considerable resources locating, acquiring and scanning books while the resulting digital copies remain inside private corporate datasets. A physical copy can disappear, a high-quality digital representation can be created, and public access to the work can remain no better than before.

That is a remarkably poor outcome from a process that could instead become a substantial programme of cultural preservation.

Why AI Companies Want Physical Books

For a model developer, most books are valuable primarily for the information printed on their pages. The binding matters little. The dust jacket matters little. If two editions contain almost identical text, their value as training material may be nearly identical even when collectors would regard the physical objects very differently.

Once the pages have been scanned and converted into machine-readable text, the physical copy may have served its commercial purpose. Destructive scanning is also fast: removing a binding allows loose pages to pass through automated document scanners rather than requiring every page to be photographed individually.

This technique is not peculiar to AI. Project Runeberg's own account of its scanning process shows a first edition of August Strindberg's Till Damaskus from 1898 being cut into loose leaves before scanning. The project noted that the book had cost only 125 kronor, while the labour required to digitise it was worth several times as much. Runeberg also says it treats genuinely rare books more carefully.

That example is useful precisely because it complicates the emotional reaction to destructive scanning. Age, rarity and preservation value are not the same thing. Destroying one copy of a mass-market paperback of which hundreds of thousands survive has little significance for cultural preservation. Destroying an unusual edition of which only a handful remain is another matter entirely.

The words are not always the only information contained in a book. Bindings, paper, illustrations, printing errors, ownership marks, handwritten annotations and marginalia can all matter to historians. An AI training pipeline may reasonably ignore most of those features. A preservation programme cannot. The sensible objection is therefore not to destructive scanning as such, but to destructive scanning without first knowing what is being destroyed.

The Long Tail Becomes Valuable

Cultural digitisation has always suffered from awkward economics. Famous books are easy to justify preserving. Obscure ones are not. A regional agricultural journal, an out-of-print engineering manual or an unsuccessful travel memoir may attract only a few readers in a decade, yet digitising it still requires somebody to locate the copy, handle it, scan it, create metadata, run optical character recognition, store the files and maintain access to them.

Projects such as Runeberg and Project Gutenberg have created enormous public value because volunteers and institutions have been willing to perform work for which there is little conventional commercial market. Their constraint has never been a shortage of books worth preserving. It has been limited money, equipment and labour.

AI alters those incentives. Model developers can have reasons to value parts of the long tail that conventional publishing and digitisation struggle to monetise. Another digital copy of Pride and Prejudice offers relatively little new textual variety to a corpus already saturated with the novel, quotations from it and writing about it. A technical handbook from 1912, a provincial memoir or a specialist monograph that never appeared electronically may contain language and information much less represented elsewhere.

The exact marginal value of any particular obscure book to a model is difficult to judge from outside the training process, and scarcity alone does not make a text useful. But books have another attraction: they often consist of sustained, edited prose rather than the mixture of duplicated pages, advertising, comments, search-engine filler and automatically generated material increasingly found across the web.

AI has therefore created a new commercial appetite for printed material that previously had little market value in digital form. This is not historically unprecedented. Google Books demonstrated decades ago that commercial interests could finance book digitisation on enormous scales. Modern AI adds a different incentive: the text itself is valuable as machine-readable training material.

One Acquisition Pipeline, Two Assets

A commercial book-scanning pipeline might look roughly like this: find books, buy them, transport them, catalogue them, remove their bindings where appropriate, scan them, extract the text, store the data and use it to train models. By the end, the company has created a valuable digital corpus. Society may have gained nothing, even if the source book was destroyed along the way.

A preservation-oriented pipeline would not be identical, but much of the expensive acquisition infrastructure would already exist. The books have been located and purchased. They have been transported to a scanning facility. Editions have been identified and prepared for digitisation. Page images have been captured.

From there, a preservation programme could retain archival-quality masters where justified, preserve bibliographic and provenance information, separate rare objects from disposable copies, deposit material with appropriate institutions and release scans publicly when rights permit.

Those additions are not free. Real digital preservation involves choices about resolution, colour, file formats, quality control, physical handling, metadata, storage and redundancy that a training pipeline may have little reason to make. Library of Congress guidance on preservation digitisation makes clear that creating a durable digital surrogate involves considerably more than producing legible OCR.

The economic claim is narrower: AI companies are already paying for many difficult first steps, particularly acquisition and basic digitisation. Building preservation onto an existing industrial scanning operation can therefore be cheaper than creating the same preservation programme from nothing. The same acquisition process can produce two assets: a proprietary training corpus and a durable cultural archive.

Copyright Complicates Access, Not Preservation

The simplest version of the proposal would be to publish every scan. Copyright makes that impossible.

The United States litigation in Bartz v. Anthropic illustrates how sharply the legal analysis can differ depending on how material was acquired and what was done with it. In a 2025 summary-judgment ruling, Anthropic's use of books to train its models was held to be fair use, and its one-to-one conversion of lawfully purchased print books into digital replacements for its internal library was also accepted as fair use. The court did not extend that conclusion to Anthropic's retention of millions of books downloaded from pirate libraries as part of a permanent corporate library. Claims relating to those pirated copies later became part of a $1.5 billion settlement that received final approval in July 2026.

None of this establishes a general rule that buying a physical book gives an AI company unrestricted rights over a digital version. Nor does it imply that a company can place copyrighted scans online merely because a particular internal use is lawful.

Public-domain material presents the clearest opportunity. When both the work and the relevant edition are free of copyright restrictions, there is a strong case for making the resulting page images, OCR and metadata available through an established archive rather than leaving them inside a private dataset. There is little reason for every AI company to create another branded digital library when institutions already exist to preserve and distribute books.

Copyrighted material requires a different arrangement. A preservation compact could create mechanisms for depositing copies with trusted archival institutions under whatever licence or statutory framework the relevant jurisdiction requires, while keeping public access restricted where copyright demands it. Works with uncertain ownership, including orphan works whose rights holders are difficult or impossible to identify, make that separation between preservation and immediate publication particularly valuable.

A book does not need to be commercially exploitable or immediately publishable online to be worth preserving.

A PDF on a Corporate Server Is Not a Cultural Archive

Digitisation by itself does not amount to preservation. A company interested primarily in extracting text may not care about colour fidelity, plates, binding details, handwritten notes or exact pagination. Its OCR may matter more than the page image, while a historian fifty years later may want almost the opposite.

A sensible preservation programme therefore needs triage rather than one expensive standard applied indiscriminately. A battered copy of a recent airport thriller with hundreds of thousands of surviving copies does not require museum-quality treatment. An unusual edition from 1820 might.

Before destructive scanning, large programmes could check bibliographic catalogues, institutional holdings, known digital copies and edition rarity. Scarce or historically significant objects could be scanned non-destructively or transferred to institutions equipped to preserve them; common books could continue through high-speed workflows.

AI itself should eventually make that triage easier. Matching editions across library catalogues, extracting bibliographic information and identifying unusually scarce holdings are exactly the kinds of classification problems modern systems can assist with. The machines may need only the text, but that does not require treating every physical copy as though nothing else about it matters.

The Public-Relations Opportunity

AI companies are trying to persuade governments and the public that their technology will create broad social benefits while simultaneously fighting disputes over web scraping, copyrighted training material, pirate libraries and compensation for writers and other creators. Against that background, buying rare books and feeding them into industrial cutters is close to a public-relations parody.

The same activity could support a very different story. Imagine an AI company reporting not merely how many books it had scanned, but how many previously undigitised public-domain works had been transferred to open libraries; how many restricted works had been placed in long-term preservation; and how many physically rare copies had been identified before destructive scanning and donated to libraries instead.

The company would still obtain the text it wanted for training, while the public could point to something concrete that existed because the training programme existed: books previously available only on paper would now survive in durable digital form.

There is a real trade-off. Public-domain does not mean commercially worthless. An obscure book may be legally free for anyone to copy while remaining expensive to locate, acquire, scan and clean. Releasing a good digital copy therefore gives competitors an asset they might otherwise have had to create themselves. The competitive advantage is not imaginary; it is simply narrower than ownership of the underlying work.

The question is whether retaining that advantage is worth more than the cultural and reputational value of releasing the scan. For an industry whose freedom to operate increasingly depends on political and public legitimacy, the answer may often be no. A company able to say, credibly, that public access to old books improved because it trained on them would have something more persuasive than another advertising campaign about responsible AI: a measurable public benefit alongside the private one.

An AI Digitisation Compact

Much of this does not need to wait for legislation. A voluntary preservation compact could begin with a handful of practical commitments:

• Check before destroying. Determine whether the book is already adequately digitised and whether the particular edition or copy appears unusually scarce.
• Match preservation effort to the material. Common modern books can pass through efficient workflows; rare or historically significant objects deserve non-destructive treatment or transfer to an appropriate institution.
• Retain archival-quality material where it matters. Training-quality OCR is not necessarily preservation-quality digitisation. When a genuinely new or significant work is scanned, preserve a master suitable for future use.
• Preserve metadata and provenance. Knowing the words in a corpus is not the same as knowing which edition produced them, when it was published or where the copy came from.
• Release public-domain scans through established institutions. Existing libraries and digitisation projects are better suited to long-term access than a corporate website created for the occasion.
• Create lawful preservation routes for restricted material. Where copyright prevents open publication, companies and archival institutions can develop deposit arrangements that preserve works without prematurely making them public.
• Publish the numbers. Report how many books were scanned, how many were already available digitally, how many public-domain works were released, how many restricted works were deposited and how many scarce physical copies were diverted from destruction.

None of this requires revealing model architecture, proprietary preprocessing methods or the complete composition of a training corpus. It turns cultural preservation from an accidental possibility into a deliberate by-product of industrial digitisation.

The controversy over book scanning is easily absorbed into the much larger argument over whether AI companies should be allowed to train on copyrighted material at all. That argument matters, but preservation does not need to wait for it. If courts and legislators ultimately permit broad training on legally acquired books, the preservation case still stands. If they require extensive licensing instead, it still stands.

The question of whether an author should be paid when a model trains on a novel is different from the question of whether an obscure public-domain book should disappear after the industry has already done the expensive work of turning it into a digital object.

Projects such as Runeberg demonstrate that old printed works can simultaneously become human-readable archives and machine-readable data. Runeberg even makes digitised textual datasets available for natural-language processing and text mining. Preservation and computational use are not inherently opposed.

The striking development is that artificial intelligence has created a new commercial appetite for parts of the printed record that conventional markets largely ignored. For decades, preservation projects have struggled to find enough money and labour to bring obscure physical books into the digital world. The AI industry has now found its own reasons to pay for much of that process.

It would be absurd to insist that companies should not digitise books simply because their motives are commercial. It would be equally absurd to build industrial-scale book-digitisation infrastructure, destroy some of the source material and then allow every resulting preservation opportunity to vanish behind corporate walls.

The industry only needs to digitise the physical book once. Where the law allows it, there is no reason the model should be the only thing allowed to remember it.