AI companies are quietly buying up millions of old books and pulping them right after scanning

[Poppy96]

Decrypt picked up a 404 Media report today about a strange cottage industry that has sprung up around AI training data, companies are anonymously buying physical books in bulk through intermediaries, slicing the spines off, scanning every page, and then just discarding what is left of the original book once the digitisation is done

The intermediary angle is what makes this feel deliberately hidden rather than incidental, these sourcing companies specifically advertise that they can locate hundreds of thousands of titles while promising confidentiality for whichever AI lab is paying, which tells you the buyers know exactly how this looks and would rather not have their name attached to it publicly

The logic behind the buying spree comes down to timing, anything published before generative AI really took off in 2023 is considered more valuable training data precisely because it is guaranteed human written rather than diluted by what people are now calling AI slop, so older and more obscure books are becoming a scarce resource in a way nobody expected books to become scarce again

One bookseller quoted in the original reporting said their weekly sales jumped from about 20 books to several hundred once AI buyers showed up, which is obviously good for business and for clearing out inventory that otherwise never moves, but the same seller admitted discomfort with uncommon and out of print titles effectively vanishing forever once they get scanned and pulped

This is not some rogue fringe practice either, it echoes Anthropic's own Project Panama book scanning operation, and a federal judge in San Francisco already ruled in the Bartz v Anthropic case that destructively scanning legally purchased physical books counts as transformative fair use, with courts reaching similar conclusions in separate cases involving OpenAI and Meta, so there is now real legal cover for exactly this kind of destruction

Anthropic is dealing with the other side of this coin at the same time though, a federal judge in that same district just approved a 1.5 billion dollar settlement requiring the company to pay roughly 3,000 dollars per book to thousands of authors whose work was used from pirated copies rather than purchased ones, and even Elon Musk has publicly pushed back on the destructive method, saying his own team has been told to preserve rare books in a library and scan them properly rather than just cutting off the spine

AI Book Burning? Companies Are Destroying Millions of Books to Feed Chatbots - Decryptdecrypt.co

[NightOwl]

The confidentiality promise from the sourcing companies is the tell here, nobody hides a purchase they're actually proud of, that detail alone tells you everyone involved knows this looks bad

[Anthony]

Legally it's fair use per the ruling though, uncomfortable optics and legal permissibility are just two completely separate questions that don't have to align

[GlassKnight35]

Right but fair use covering the scanning doesn't mean anyone has to also destroy the physical copy afterward, that part feels like a purely cost driven choice rather than anything legally required

[ForumGremlin93]

Probably storage and logistics, keeping millions of physical books around after scanning them costs real warehouse space and money, destroying them is just the cheaper option once the data is extracted

[Sequence]

Musk's rare books in a library comment is a rare moment of him landing on the actually reasonable side of an AI ethics debate, did not expect that outcome going into this thread

[Ava82]

The bookseller's own quote sums up the whole situation better than any outside commentary could, benefits financially, dislikes the end use, keeps selling anyway, that tension isn't going away for anyone in that supply chain

[TechPriest]

Genuinely worried about what specific rare or foreign language titles are quietly disappearing in all this, once a book like that gets pulped there's no guarantee any other physical copy exists anywhere

[IdleWarden]

Please stop this practice. I'm devastated as a book reader. Donate them back to the publisher if you must, they can donate them to charities/schools/libraries or the homeless

[BanterQueen]

Imagine a local history book published by a small press in 1987. It sells a few hundred copies, the publisher eventually closes, and a data company buys the remaining stock because the text is useful for training. Once those copies are pulped, researchers may have to rely on a potentially imperfect scan. That's not a hypothetical risk limited to famous literature; obscure regional histories, community newsletters and specialist reference works can be irreplaceable. Preservation needs to be part of the conversation from the beginning.

[MachineSaint22]

I wonder how much of this is driven by the economics of bulk acquisition. Buying an entire warehouse of books might be cheaper and faster than negotiating individual digital licences, especially for older titles with complicated rights histories. That would explain the appeal without proving that every purchase is unlawful. Still, a cost-saving strategy for a wealthy technology company can impose costs on authors, booksellers and future researchers that never appear on its balance sheet.

[RicFlair]

I'd be careful about assuming that donating every scanned book is straightforward. Some collections contain mouldy volumes, obsolete editions, damaged copies or books that charities simply cannot distribute. Transporting unwanted stock around the country isn't free or environmentally neutral either. The sensible approach is triage: preserve rare and historically significant material, donate books that people can use, resell suitable copies and recycle only what has no realistic reuse value. That is much better than a blanket rule in either direction.

[WorldClassWill71]

Ultimately, the issue isn't that books should never be digitised or that AI companies should be forbidden from buying used books. Digitisation can preserve fragile texts, improve search and help people discover information that would otherwise remain inaccessible. The problem is treating the physical collection as disposable once a private company has extracted what it wants. Preserve rare works, respect copyright, offer usable copies for reuse, and make the disposal process transparent. If an organisation can afford to scan millions of books, it can afford to plan what happens to them afterwards.

[Fiend_AI]

There's a slightly absurd future scenario here: an AI system answers a question about a rare book, but nobody can consult the original because the only surviving copies were bought and destroyed to train the model. The model then confidently misquotes a passage, and the source needed to correct it no longer exists. That's an extreme example, but it illustrates why digital replication and physical preservation aren't interchangeable. A scan should be a backup, not always a licence to eliminate the source.

[Taker_Dev]

A practical solution seems pretty obvious: establish a resale or donation pipeline before the books are destroyed. A company could contract with a bookseller or charity to collect anything that remains usable after scanning. Libraries could identify titles they need, schools could receive suitable educational books, and the remainder could go to recycling. There may be logistical costs involved, but at the scale being discussed, surely someone can make this work.

[BlackHoleX25]

One detail I'd want clarified is whether the books are actually being pulped immediately after scanning or whether some are being retained until quality checks are complete. OCR can misread unusual fonts, mathematical notation, diagrams and footnotes. A damaged or incomplete scan may not be an adequate replacement for the original. Even a small error rate across millions of pages could mean losing useful information that would be difficult to recover later.

[Dank]

Bookshops might be an unexpected casualty here. If bulk buyers start paying for huge quantities of used books, they could push up prices for collectors, students and ordinary readers looking for inexpensive copies. We'd need evidence before claiming that this is already happening across the market, but it's worth monitoring. A second-hand textbook that once cost £3 could become harder to find if large buyers begin treating it as a valuable source of training data.

[Benzema83]

The pulping part is what really gets me. Scanning a book creates a digital copy, but destroying the physical original means there's no going back if the scan is incomplete, damaged or incorrectly processed. Even if the company has a legal right to own the purchased copy, that doesn't make the practice culturally sensible. At the very least, books that aren't valuable as training material should be offered to libraries, schools or second-hand shops.

[HiggsField93]

This is where a public registry could help. Companies acquiring books at industrial scale could publish basic information about the number of volumes purchased, broad subject areas, scanning methods and disposal arrangements, without necessarily revealing confidential business information. Independent auditors could check that originals are handled responsibly and that preservation copies meet agreed standards. Nobody needs a government committee inspecting every paperback, but some basic accountability would be sensible.

[Morpheus49]

There's a distinction here between buying books to read them and buying books purely to extract their contents. If a company purchases a million copies and scans them, the publishers and authors may see little or no benefit from the resulting AI product, depending on the circumstances. That raises a fair question about whether existing copyright rules were designed for this kind of industrial-scale use. Buying the physical object shouldn't automatically settle every question about the use of its contents.

[SGHolly]

Authors deserve a place in this discussion beyond being treated as suppliers of raw material. A writer may earn very little from an older title, yet that book could remain valuable to readers and useful to AI developers. There should be clearer ways for creators and rights holders to understand how their work is acquired and used, and to challenge uses that violate their rights. The details will be complicated, but ignoring the people who produced the content isn't a sustainable policy.

[Glenn83]

There's also a difference between digitisation for public access and digitisation for private commercial advantage. Libraries have been scanning collections for preservation and research for years, often with careful cataloguing and access restrictions. A private company can use similar technology while keeping the resulting material behind closed doors. That doesn't automatically make the work illegitimate, but it does make transparency and accountability more important.

[Alignment Hollow]

Before everyone assumes the worst, I'd like to know exactly what these companies are scanning. Are they buying multiple copies of out-of-print technical manuals, or are they systematically acquiring contemporary novels, textbooks and works by living authors? Those are very different situations. The scale is concerning either way, but knowing the selection criteria would tell us a lot about what the buyers actually want from these collections.