Cut Apart for AI Training: Millions of Books and No Record Kept #
The email that arrived at an antiquarian bookshop in Haarlem in the summer of 2026 listed 3,001 ISBNs. The bookseller, Pieter de Vries, assumed it was fraud. “Very unusual,” he told Fortune. “That's why I regarded the request as spam or phishing.” He deals in old and rare books, he explained, the kind that are expensive and exclusive, and nobody orders three thousand of those at once. When he later learned what such requests were widely believed to be for, his reaction was two words: “I was shocked!”
The detail that matters most about that list is the one least likely to survive a retelling. According to Fortune's account, the titles were mostly published in 2020 and 2021, and came from academic houses — Emerald, Elsevier, Wiley, Routledge, Oxford University Press. They were not incunabula. They were recent monographs, the sort of thing a university library buys and a secondhand dealer eventually inherits. The requesting party was 2077AI, which presents itself publicly as a Singapore-based non-profit open-data foundation building open datasets and benchmarks. It did not respond to Fortune's requests for comment.
This is the gap the story has been falling through since late July. There is a real, documented, court-evidenced programme of industrial book destruction carried out by a major AI laboratory. There is also a viral narrative about the last surviving copies of eighteenth-century texts being fed into shredders, which has almost no verifiable sourcing behind it. The two have fused into a single story, and the fusion is doing damage to both — it makes the real scandal easier to dismiss as hysteria, and it makes the hysteria harder to correct without appearing to defend the scandal.
So it is worth separating them carefully, because the actual cultural cost is not quite where the outrage has located it, and it is arguably worse.
What the Court Documents Actually Established #
The foundational reporting is not from July 2026. It is from January.
On 27 January 2026, The Washington Post published an account, drawn from documents unsealed in copyright litigation, of an internal Anthropic initiative called Project Panama. An internal planning document put the objective without euphemism: “Project Panama is our effort to destructively scan all the books in the world.” The same document instructed staff to keep it quiet. “We don't want it to be known that we are working on this.”
The mechanics, as described in the unsealed material and reported subsequently by outlets including IBTimes, were industrial rather than exotic. Anthropic hired Tom Turvey, formerly head of partnerships at Google Books, in February 2024, tasked according to internal documents with obtaining “all the books in the world”. Purchased volumes went to vendors who used hydraulic cutting machines to remove the spines, fed the resulting loose sheets through high-speed production scanners, and arranged for a recycling company to collect what was left. Vendor proposals reported by The Washington Post referenced converting somewhere between 500,000 and two million books over a six-month period. The company spent tens of millions of dollars on the stock.
None of this was a leak in the conventional sense. It emerged because the authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson sued, and discovery did what discovery does.
The reason the story detonated again in late July 2026 is a matter of coincidence and algorithm. On 20 July, Judge Araceli Martínez-Olguín of the Northern District of California granted final approval to the $1.5 billion settlement of the piracy claims in that case — a figure the Authors Guild describes as delivering roughly $3,000 per work across an estimated 500,000 works. Six days later, on 26 July, an account on X posted a compressed summary of the whole affair. Elon Musk replied to it the following day. Within seventy-two hours, the story had been rewritten by dozens of outlets, many of them working from each other rather than from the January reporting or the court record.
That is how a January investigation into a 2024 programme became an August emergency.
The Ruling That Made Destruction the Prudent Choice #
The legal architecture here is the part almost every retelling gets backwards, and it is the part that should worry people most.
On 23 June 2025, Senior District Judge William Alsup issued a partial summary judgment order that split Anthropic's conduct into three categories with three different answers. Training large language models on the books was, in his phrasing, exceedingly transformative, and therefore fair use. Down more than seven million books from Library Genesis and the Pirate Library Mirror to build a permanent central library was not fair use, and was the conduct that produced the settlement. And the third category — the destructive scanning of lawfully purchased print copies — was fair use as a format change.
Alsup's reasoning on that third point deserves quoting rather than paraphrasing, because its logic has consequences. Every purchased print copy, he wrote, was copied in order to save storage space and enable searchability as a digital copy. The print original was destroyed. One replaced the other.
Read that sequence again. The destruction is not incidental to the legal finding. It is load-bearing. What made the format shift defensible was precisely that no additional copy came into existence: the company that bought one book ended up with one book, in a different substrate. A firm that scanned non-destructively and kept the physical volume would have ended the process holding two copies where it had lawfully acquired one, and would have needed a different and less certain argument.
The first-sale doctrine does the rest of the work. Buy a book and you may burn it. Nothing in copyright law obliges a purchaser to preserve the object, and nothing obliges a purchaser to disclose to the seller what they intend to do with it.
The result is a legal regime in which the cheapest method of digitisation is also the safest method, and the method that preserves the artefact carries both a higher cost and a marginally higher legal risk. Whatever one thinks about the cultural stakes, that is a badly designed incentive. It is not that the law permits destruction. It is that the law, as applied, quietly rewards it.
What a Guillotine Removes #
Destructive scanning is not a sinister technology. It is a boring one, and it is used every day by document bureaux digitising insurance files.
A hydraulic or guillotine cutter shears the bound edge off the text block, converting a book into a stack of loose leaves. Those leaves go through an automatic document feeder at rates that commercial operators quote in the thousands of pages per hour. The alternative — planetary or overhead scanning, where a bound volume rests in a cradle and a camera photographs each opening while a human or a robot turns pages — is slower, requires more expensive equipment, and produces images that need more correction for curvature and gutter shadow. Published price guides for digitisation services put the range across both methods at roughly ten cents to a pound and a half per page, with the destructive route sitting near the bottom of that band and bound-volume capture near the top.
So the trade is speed and unit cost against the survival of the object. On a two-million-book programme, that difference is real money. Against a training budget measured in billions, it is a rounding error — which is the substance of Musk's argument, and it is a good one.
What the guillotine takes is not the text. The text survives, in the sense that the words can be extracted. What it takes is everything about the copy that is not the text, and this is where rare-book librarians and bibliographers part company with engineers.
The guidance published by IFLA's Rare Books and Special Collections Section on planning digitisation projects is explicit that the physical object carries evidence the surrogate cannot. Provenance evidence — bookplates, ownership signatures, library stamps, booksellers' pencil codes, auction records tied to a specific copy — establishes where a book has been and who read it. Marginalia record what readers thought while they were reading, and constitute a distinct field of scholarship in their own right; the Early Modern Marginalia Research Network exists because annotations in the gutter of one particular copy are historical evidence of a kind nothing else supplies. Binding structures, sewing patterns, paper stock and watermarks date and locate production. Collation — the physical gathering structure — is how bibliographers distinguish issues and states of the same nominal edition, and how they detect made-up copies, cancels and substituted leaves.
A guillotine destroys collation evidence absolutely. It destroys the binding. It destroys the relationship between a specific annotation and a specific gathering. And a scan optimised for text extraction will not preserve any of it as a surrogate, because it is not trying to: training-oriented capture wants clean, legible, machine-readable characters, not colour-managed high-resolution page images with calibrated targets and structural metadata. There is no public evidence that Anthropic's scanning produced preservation-grade image files, and considerable evidence that it did not need to.
This distinction matters for a reason that goes beyond sentiment. A preservation scan is a surrogate that a future researcher can interrogate. A training scan is an extraction. When the extraction is done and the sheets have gone to the recycler, what remains is text inside a corporate storage system, and there is no public record of the correspondence between that text and the objects it came from.
Whether Anything Unique Was Actually Destroyed #
Here is where the story needs discipline, because the most emotionally potent claim in circulation is also the least supported.
The claim, repeated across aggregator sites and social posts since late July, is that the last three copies of an eighteenth-century botanical text were shredded. I could not find a primary source for it. It does not appear in The Washington Post's January reporting, in 404 Media's investigation, in the Fortune piece, or in the Alsup order. It appears in commentary blogs and in downstream summaries, several of which cite each other. Until someone produces a title, an ISBN or accession record, and a holdings check, it should be treated as unverified, and outlets repeating it as fact are not helping.
The companies named have all denied the strong version. An Anthropic spokesperson has stated that none of its data acquisition programmes buys and destroys “rare” or “antiquarian” books, that it procures from mainstream commercial markets, and that it never bought from Zoom Books, the Canadian firm at the centre of the Australian and European bookseller complaints. Zoom Books, responding in May 2026 to accusations published by the Spanish outlet Demócrata, said that it acquires secondhand books and resells them intact, and that it does not digitise or destroy books for the purpose of training AI models, calling contrary claims inaccurate; it added, The Atlantic reported, that it did not to its knowledge sell books for those purposes. ISBNdb, whose marketing pages 404 Media exposed on 21 July, removed those pages nine days later and posted a denial stating that it has never purchased, scanned or sold a book, that it does not train AI models, and that the page was a test of market interest for a service never launched.
Zoom's qualifier matters: the 2024 destruction is documented in court records; this year's bulk orders are not attributed. Investigating for The Atlantic in early August, Alex Reisner found the surge real but the end buyer unconfirmed: some Zoom orders shipped to PrepFort, an Amazon and Walmart logistics firm that declined to discuss its clients. Reisner judged it very possible the books were bound for AI companies, but not certain. A Houston bookseller, Charlie Becker, read his own orders as Amazon arbitrage: underpriced books bought to relist at a markup.
One strand has since been traced: on 17 August, 404 Media reported that an AirTag hidden in a bulk order had ended at an Amazon facility in Las Vegas where, an anonymous former worker said, spines are cut and pages scanned. Amazon said only that it “purchases books through commercial channels to help develop and improve the products and services our customers use.” That links one order to one company, and no more.
Denials are not proof. But nor is their absence, and there is currently no documented instance of a specific unique or near-unique title being destroyed by an AI training programme. What is documented is the destruction of millions of ordinary in-copyright books, and bulk purchase requests for recent academic monographs.
Anthropic's denial is less tidy than it sounds. Unsealed emails reviewed by Snopes show Turvey telling Wonder Book, a bulk bookseller, in August 2024 that “less common books are a great place to start.” Asked to clarify, Anthropic distinguished “rare and valuable” books from harder-to-find professional and reference titles. That distinction concerns price, not how many copies survive.
The counterweight comes from library science, and it complicates both sides. Research by OCLC on what its authors call the collective collection has repeatedly found that scarcity in print holdings is the norm rather than the exception. In one analysis of the Committee on Institutional Cooperation's shared print collection, roughly three-quarters of titles were held by three or fewer member libraries. In work on the collective collection of Research Libraries UK, close to ninety per cent of print book publications were held by fewer than five libraries in the group. Across large mega-regional aggregations, at least three-quarters of titles sat in five or fewer institutions.
Two conclusions follow, and they point in opposite directions. The first is that “rare” is close to meaningless as a category boundary, because thin holdings are ordinary. The second is that thin holdings are precisely what makes bulk destruction risky: if most titles survive in only a handful of institutional copies, then the secondhand market is not a reservoir of redundancy. It is a thin layer over a thin layer.
Deposit systems provide a partial floor. Under the Legal Deposit Libraries Act 2003, a copy of every book published in the United Kingdom must be deposited with the British Library, with five further libraries entitled to request copies. Comparable regimes operate in most publishing nations. Anything published commercially in a country with functioning legal deposit almost certainly survives somewhere in an institutional collection, regardless of what happens to the secondhand stock.
The categories that fall outside that floor are the ones nobody is checking: self-published and small-press work, regional and community publications, non-commercial technical and trade literature, foreign-language material published in jurisdictions with weak deposit enforcement, and the enormous grey zone of items that were deposited but never catalogued into any union catalogue searchable today.
Which brings us to the actual finding, and it is not the one either camp wants. Nobody knows whether a unique copy was destroyed, because nobody kept a manifest. There is no published, title-level record of what went through the cutters. No holdings check was performed. No threshold was set. The absence of that record is not a side issue; it is the centre of the case. A programme that destroys books at industrial scale without logging what it destroys has made its own claim of innocence unfalsifiable, and has simultaneously made the accusation against it unfalsifiable too. That is a choice, and it was Anthropic's to make differently.
Whether the Paper Was Really Necessary #
The demand-side justification, as ISBNdb marketed it to prospective clients, is elegant. “The world's best AI training data is sitting on a shelf,” the pitch read, describing books as curated, peer-reviewed, domain-specific human knowledge, structured in a way no web crawl can replicate — dense, edited, authoritative. And the pages carried a candid aside that 404 Media quoted and which has since done more reputational damage than anything else in the story: the optics problem is real, because “AI company destroys two million books” is not a headline that generates sympathy.
The underlying scientific worry is genuine. Ilia Shumailov and colleagues published a paper in Nature in July 2024 showing that generative models trained recursively on their own outputs degrade, losing the tails of the distribution and converging on blandness — the phenomenon popularised as model collapse. The paper received an author correction in Nature in 2025 and remains widely cited. If the open web is progressively filling with synthetic text, then pre-2022 print is a supply of guaranteed-human material, and its scarcity value rises with every model release.
But the literature is less apocalyptic than the marketing. A response by Ali Borji argued that the reported behaviour is a general statistical consequence of repeatedly fitting a distribution and resampling from it, rather than a novel property of language models. More consequentially, Matthias Gerstgrasser, Rylan Schaeffer and co-authors demonstrated in 2024 that collapse follows from replacing real data with each generation's synthetic output, and that accumulating synthetic data alongside the original real data avoids collapse — a result they showed across model sizes, architectures and hyperparameters, and reproduced in diffusion models and variational autoencoders. Nobody in the industry is deleting their pre-2022 corpora. The realistic scenario is accumulation, which is the regime in which collapse does not occur.
Scarcity of human text is also a softer constraint than the pitch implies. Epoch AI's analysis puts the effective stock of human-generated public text at around 300 trillion tokens, with a wide confidence interval, and projects full utilisation somewhere between 2026 and 2032 — an estimate that has already been pushed back once by improvements in data efficiency and repeated-epoch training.
Set the numbers side by side and the arithmetic is instructive. Take the upper end of the vendor proposals reported by The Washington Post — two million books — and assume a generous average of around 100,000 words per volume. That is on the order of 260 billion tokens: roughly a tenth of one per cent of Epoch's central estimate of the total human text stock. The case for print therefore cannot rest on volume. It rests entirely on a claim about quality weighting — that an edited monograph is worth some large multiple of an equivalent quantity of web text — and no laboratory has published the numbers that would let anyone outside evaluate that multiple.
This is the point at which the trade-off in the original question becomes difficult to state honestly. Is a unique book worth an incremental improvement in model performance? Nobody has quantified the improvement. The industry is asking the public to accept an irreversible cost against an unspecified and unaudited benefit, which is not a trade-off so much as a request for trust.
And the trade-off is in any case a false binary, because destruction was never required to obtain the text. Google Books digitised more than forty million volumes using predominantly non-destructive camera rigs. The Internet Archive scans non-destructively and then, rather than pulping, maintains a physical archive in Richmond, California, storing volumes in modified shipping containers held at controlled temperature and around thirty per cent relative humidity. Both organisations solved the engineering problem of mass digitisation without solving it by disposal. The difference is cost per unit and, in Anthropic's case, the legal convenience of ending up with exactly one copy.
What the Booksellers Are Objecting To #
The booksellers' complaint is not a legal complaint, and pretending otherwise weakens it.
In Australia, Guardian reporting recorded a pattern that dealers found unmistakable once they saw it. John Sainsbury of Sainsburys Books in Melbourne described a run of orders from Zoom Books for titles with no relationship to one another — engineering, cycling, poetry, local history — placed with apparent indifference to price. Delfina Manor of Good Reading Secondhand Books in Benalla, Victoria received an order in May large enough that fulfilling it became an ordeal: “It was just driving me nuts.” A German dealer reported orders arriving between three and five in the morning. Booksellers quoted by The Guardian described the practice of cutting spines and discarding the remains as horrific. The Telegraph's Chris Stokel-Walker headlined his 30 July piece with the phrase cultural barbarism. A seller quoted anonymously by 404 Media captured the ambivalence better than either: the buying benefits him financially, but “I don't like the end-use, and I don't like that uncommon books are being pulped.”
Under first sale, none of this gives rise to a cause of action. A buyer may destroy what they own and need not explain themselves. If the objection were simply that destruction is wrong, it would prove too much: the book trade already pulps enormous quantities of unsold and unsellable stock, and remainder destruction is a routine part of publishing economics.
The objection rests on three narrower things, and they are stronger for being narrow.
The first is deliberate opacity. ISBNdb's withdrawn marketing did not merely offer sourcing; according to 404 Media's reporting it offered anonymity, positioning buyer confidentiality as a feature and naming the reputational reason. Anthropic's own planning document recorded that it did not want the work known. Concealment of end use is not fraud, but it is a considered decision to prevent the counterparty from exercising a judgement they would otherwise have exercised. Booksellers are not claiming a right to control what buyers do. They are claiming that they were placed on the wrong side of an information asymmetry that was engineered on purpose. Routed through confidential intermediaries, the 2026 orders leave sellers unable to tell whether a parcel ends in a scanner, an Amazon relisting or a recycler.
The second is role morality. Antiquarian and secondhand dealers do not generally think of themselves as retailers of interchangeable units. The trade's self-conception is custodial: a book passes through your hands on its way to the next reader, and part of the job is judging where it should go. That is a professional identity, not a legal entitlement, but it is the reason the reaction has been visceral rather than commercial.
The third is market structure. Price-insensitive bulk buying in a thin market does not merely remove copies; it distorts the pricing signals dealers use to allocate scarce stock, and it strips the mid-list and specialist tiers fastest, because those are where supply is thinnest and the buyer's indifference to price bites hardest. If a programme wants two million books and does not care which two million, the books it takes are the ones nobody else was competing for — which is another way of describing the scarce ones.
The Precedent That Should Have Settled This #
None of this is unprecedented, and the precedent is close enough that its absence from most of the coverage is conspicuous.
In April 2001, Nicholson Baker published Double Fold: Libraries and the Assault on Paper, expanded from a New Yorker piece of July 2000. Its subject was the mid-century programme under which American research libraries, including the Library of Congress, microfilmed newspapers and books and then discarded the originals, on the reasoning that wood-pulp paper was embrittling towards inevitable disintegration. Baker argued that the brittleness case was overstated, that the space savings were the real motive, and that irreplaceable runs of newspapers had been destroyed to produce surrogates that were often poorly made, badly filmed, and stripped of the colour supplements and physical evidence the originals carried.
The book was contested immediately — Richard Cox's response for the Society of American Archivists is the best-known rebuttal — and Baker conceded in his preface that it was not an impartial piece of reporting. Some of his specific charges were unfair to librarians working under genuine preservation pressures and real budget constraints.
But the structural lesson survived the argument, and it is precisely applicable. Surrogate-making organisations systematically underestimate what the original carries, because they define value in terms of the thing they are capturing. The loss is invisible until someone needs the original for a purpose the surrogate-maker did not anticipate. And by the time that purpose arises, the decision is unreviewable.
Apply that to a training scan. The purpose is text extraction. Everything defined as noise — the binding, the annotation, the paper, the collation, the ownership marks — is discarded twice: once when the capture is specified, and again when the sheets go to the recycler. If a future researcher, or for that matter a court in some later case weighing Alsup's reasoning, wanted to examine the physical objects, there is nothing to examine.
Musk, and the Limits of a Pledge #
On 27 July 2026, replying to the X post that had propelled the story, Elon Musk wrote: “I've asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning.”
Two corrections are warranted, one to the sceptics and one to the enthusiasts.
To the sceptics: SpaceXAI is a real entity. CNBC reported that SpaceX acquired xAI in an all-stock transaction announced on 2 February 2026, valuing the combined business at $1.25 trillion, folding xAI, Grok and the X platform into SpaceX; the merged company listed on Nasdaq in June 2026, and the xAI name was retired in favour of SpaceXAI in July. Dismissing the pledge on the grounds that no such company exists is wrong.
To the enthusiasts: the pledge is close to unfalsifiable. It does not define “rare”. It specifies no holdings threshold, no catalogue against which rarity would be assessed, no verification, no audit, no reporting, and no consequence for non-compliance. It is a sentence on a social platform. Musk's associated observation — that the extra cost of careful scanning is trivial against budgets in the billions, and that slower capture still yields the training data while the book survives — is entirely correct and is the strongest argument anyone in the industry has made. But an argument is not a policy, and no published policy has followed.
That this is nonetheless the most substantive public commitment any frontier laboratory has made on the question tells you where the floor currently sits.
What a Serious Response Would Require #
The useful question is not whether shredding is barbarism. It is what a competent regime would look like, given that mass digitisation of print is going to continue and that most of it is unobjectionable.
Six measures would do most of the work, and none of them is technically hard.
A destruction manifest, published at title level. Every volume destroyed, logged by ISBN or bibliographic identifier, with the log made public or lodged with a deposit library. This is the cheapest measure and the most important, because it converts an unanswerable argument into a checkable one. Its absence is why nobody can say whether the eighteenth-century botanical text ever existed.
A holdings check before the cutter. Union catalogues exist. WorldCat, Library Hub Discover and national union catalogues can be queried programmatically. A rule that any title falling below a stated threshold of recorded holdings is routed away from destruction would be trivial to implement at the scale these programmes operate, and would cost a fraction of a per cent of the acquisition budget.
Deposit after scanning. Where a title clears the threshold and is destroyed anyway, the alternative for flagged items should be that the intact volume is offered to a deposit or research library, with the acquiring company funding carriage. Where the volume has already been cut, the loose text block should be boxed and offered rather than recycled: a guillotined block is not an artefact, but it is not nothing, and it preserves the leaves, the annotations and the paper.
Preservation-grade capture where destruction occurs. If a copy is going to be destroyed, the images taken from it should meet archival imaging standards rather than extraction standards, with the page images retained and escrowed rather than discarded after text extraction. The marginal cost of capturing at higher resolution during a pass you are making anyway is small. The marginal value, given that the original will not exist afterwards, is large.
Disclosure at the point of sale. A simple end-use declaration to the seller — this order is for digitisation and the volumes will be destroyed — resolves the consent objection entirely. Sellers who object can decline or price accordingly. Sellers who do not care can proceed. The only thing disclosure costs is the anonymity that ISBNdb was explicitly marketing.
A regulatory hook that already exists. Article 53 of the EU AI Act obliges providers of general-purpose AI models to publish a sufficiently detailed summary of training content; the Commission published its mandatory template on 24 July 2025, with obligations applying from 2 August 2025 and a 2027 deadline for models already on the market. The Commission has been able to enforce them, with fines, since 2 August 2026, and the Digital Omnibus did not delay them. The template covers data origin and curation. It would take no new legislation, only interpretive guidance, to establish that acquisition provenance — including whether physical originals were destroyed — falls within scope. That is the single most available lever anyone concerned about this has.
The Record of What Was Lost #
So: does the permanent loss of a book that may exist nowhere else constitute an acceptable price for incrementally improving a model?
The question, posed that way, cannot be answered, and it is worth being clear about why. The loss has not been demonstrated — no unique title has been identified, and the most-repeated claim to the contrary has no traceable source. The improvement has not been quantified — no laboratory has published what two million books bought them. And the trade is false in any case, because Google and the Internet Archive have both demonstrated that the text can be obtained at scale without the destruction, and Musk is right that the cost differential is negligible against these budgets.
What can be said is narrower and, I think, more damning than the viral version.
An organisation set out to destructively scan, in its own words, all the books in the world. It recorded in writing that it did not want this known. It destroyed millions of volumes without checking, at any point, how many other copies of any of them existed. It kept no public record of what it destroyed. It captured those volumes to a standard suited to text extraction rather than preservation, ensuring that even the digital residue is a diminished thing. And the legal reasoning that blessed the arrangement did so partly because the physical copy was destroyed — meaning the destruction was not a regrettable byproduct but a component of the defence.
The cultural cost is therefore not primarily the books. It is the erasure of the record of the books: the fact that this transaction between the physical past and the computational present was conducted without a ledger, and that we are now reduced to arguing about botanical texts nobody can name because the only people who could settle the question chose not to write anything down.
Nicholson Baker's complaint about the microfilmers was never really that they preferred film to paper. It was that they made an irreversible decision on behalf of everyone who came afterwards, using a definition of value that suited their own project, and that they did it without keeping account. A quarter of a century later, with better machines and vastly more money, the same decision has been made the same way. The remedy has not changed either. Check the holdings. Keep the manifest. Offer the copy. Tell the seller.
None of that would have cost anything meaningful. That it was not done is the answer to the question.
References #
- Fortune, “This Dutch bookseller thought a request for 3,000 copies was 'spam or phishing.' Instead, AI companies are scanning and destroying books to train AI”, 31 July 2026. https://fortune.com/2026/07/31/dutch-bookseller-ai-spam-phishing-3000-book-copies-scan-destroy/
- The Washington Post, “Anthropic 'destructively' scanned millions of books to build Claude”, 27 January 2026. https://www.washingtonpost.com/technology/2026/01/27/anthropic-ai-scan-destroy-books/
- Emanuel Maiberg, 404 Media, “AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop”, 21 July 2026. https://www.404media.co/ai-companies-are-buying-tons-of-old-books-because-theyre-free-of-ai-slop/
- 404 Media, “Company Offering Printed Books to Train AI Stops After 404 Media Coverage”, 30 July 2026. https://www.404media.co/ai-company-training-scanning-books-database-isbndb/
- IBTimes UK, “Inside Project Panama, Anthropic's Secret Effort To Scan and Shred the World's Books”, 2026. https://www.ibtimes.co.uk/anthropic-secret-book-scanning-operation-1811155
- Chris Stokel-Walker, The Telegraph, “Cultural barbarism: How AI companies are destroying the world's books”, 30 July 2026 (paywalled; text and citation recorded at Kip Currier, Management and Leadership, 30 July 2026). http://kipcurriermanagementleadership.blogspot.com/2026/07/cultural-barbarism-how-ai-companies-are.html
- Outlook India, “Explained: Why Australian Booksellers Are Raising Alarm Over AI's Book Hunt”, 2 August 2026. https://www.outlookindia.com/international/explained-why-australian-booksellers-are-raising-alarm-over-ais-book-hunt
- Publishers Lunch, “Zoom Books Responds To Accusations of Selling Books For AI Scraping”, 27 May 2026. https://lunch.publishersmarketplace.com/2026/05/zoom-books-responds-to-accusations-of-selling-books-for-ai-scraping/
- United States District Court, Northern District of California, “Order on Fair Use, Bartz v. Anthropic PBC”, 23 June 2025. https://copyrightalliance.org/wp-content/uploads/2025/06/Bartz-v.-Anthropic-Order.pdf
- The Authors Guild, “Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement”, July 2026. https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/
- Elon Musk, post on X replying to @HedgieMarkets, 27 July 2026. I’ve asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning — Elon Musk (@elonmusk)July 27, 2026
- CNBC, “Musk's xAI, SpaceX combo is the biggest merger of all time, valued at $1.25 trillion”, 3 February 2026. https://www.cnbc.com/2026/02/03/musk-xai-spacex-biggest-merger-ever.html
- Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson and Yarin Gal, “AI models collapse when trained on recursively generated data”, Nature, 24 July 2024. https://www.nature.com/articles/s41586-024-07566-y
- Ali Borji, “A Note on Shumailov et al. (2024): 'AI Models Collapse When Trained on Recursively Generated Data'“, arXiv:2410.12954, 2024. https://arxiv.org/abs/2410.12954
- Matthias Gerstgrasser, Rylan Schaeffer, Apratim Dey, Rafael Rafailov and colleagues, “Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data”, arXiv:2404.01413, 1 April 2024. https://arxiv.org/abs/2404.01413
- Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim and Marius Hobbhahn, “Will we run out of data? Limits of LLM scaling based on human-generated data”, Epoch AI, 2024. https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
- IFLA Rare Books and Special Collections Section, “Guidelines for Planning the Digitization of Rare Book and Manuscript Collections”, September 2014. https://www.ifla.org/files/assets/rare-books-and-manuscripts/rbms-guidelines/guidelines-for-planning-digitization.pdf
- Brian Lavoie, “Reflections on Collective Collections”, College & Research Libraries / OCLC Research preprint, 2020. https://www.oclc.org/content/dam/research/publications/2020/oclcresearch-reflections-on-collective-collections-preprint.pdf
- Legal Deposit Libraries Act 2003, UK Public General Acts. https://www.legislation.gov.uk/ukpga/2003/28
- Nicholson Baker, Double Fold: Libraries and the Assault on Paper , Random House, April 2001.https://en.wikipedia.org/wiki/Double_Fold
- Richard J. Cox, “Don't Fold Up: Responding to Nicholson Baker's Double Fold”, Society of American Archivists, 2001. https://www2.archivists.org/news/2001/dont-fold-up-responding-to-nicholson-bakers-double-fold-by-richard-j-cox
- Alex Reisner, The Atlantic, “Someone Is Mysteriously Snapping Up Used Books Around the World”, 5 August 2026. https://www.theatlantic.com/technology/2026/08/ai-companies-buying-used-books-for-data/688167/
- Emanuel Maiberg, 404 Media, “We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility”, 17 August 2026. https://www.404media.co/we-tracked-a-shipment-of-rare-books-it-ended-at-an-amazon-ai-training-facility/
- Anna Rascouët-Paz, Snopes, “Are AI companies scanning and destroying millions of books, including rare titles?”, 3 August 2026. https://www.snopes.com/fact-check/ai-companies-destroying-rare-books/
- Mayer Brown, “EU AI Act News: Digital Omnibus on AI, New Guidance on Risk Classification, GPAI, and Transparency Obligations”, 30 July 2026. https://www.mayerbrown.com/en/insights/publications/2026/07/eu-ai-act-news-digital-omnibus-on-ai-new-guidance-on-risk-classification-gpai-and-transparency-obligations
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)
Listen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)