Anthropic’s mass-digitization pipeline has not been shown to have destroyed rare or unique books. But it exposes a perverse incentive: discarding the physical copy can cut costs while strengthening the argument that the digital file merely replaced it.

To turn a paper book into data, Anthropic’s contractors removed the binding, cut the pages to size, scanned them, and discarded the paper original.

The court record describes the process in concrete terms. Anthropic spent millions of dollars buying millions of print books, many of them used. Each book became a PDF containing page images and machine-readable text. Those files entered an internal research library. Only some sets and subsets were later selected for particular model-training mixes; the record does not establish that every scanned book trained a version of Claude.

The remarkable part is not only that this process was efficient. In a US copyright case, destroying the paper copy also strengthened the defense.

The boundaries matter.

The public court record reviewed here does not identify a rare, unique, or last-surviving copy among the books destroyed; it is not an item-level audit of every volume. Public filings concerning OpenAI establish its acquisition and use of digital LibGen-derived Books1 and Books2 datasets, not a comparable operation involving purchased paper books. Elon Musk later posted that he had asked “the SpaceXAI team” to place rare books in a library and use the more laborious method of scanning without cutting the spines. That is his wording and his account of an instruction about rare books, not an independently audited policy or evidence of what another company destroyed. On August 4, an Anthropic spokesperson publicly denied that its data-purchasing project bought and destroyed rare or antique books. The statement narrows the dispute but does not supply an item-level record outsiders can audit. Anthropic response

As of August 6, 2026, this review did not identify a complete, publicly reviewable item-level catalog, an independent audit report, or a public-preservation arrangement for Anthropic’s purchased books. That does not prove Anthropic had no internal screening. It means outsiders cannot verify what the screening was or what it found.

That is the real question:

When the company that gains the data also saves on scanning and storage—and may gain a better litigation narrative—who must prove that millions of irreversible disposal decisions were safe?

The objection is not that every mass-market paperback must be treated as a museum object. It is that the party that profits from destruction should not be the sole judge of whether a particular copy was expendable.

Copyright did not order the destruction. It changed the price of the alternatives.

Buying a book gives you ownership of that physical object. Under Section 109 of the US Copyright Act, the owner of a lawfully made copy may sell or otherwise dispose of that copy. Legislative history expressly includes destroying it.

But ownership of the object is not ownership of the copyright in the work. Section 202 makes that distinction explicit. Scanning the entire book and fixing it in a new digital file is still an act of reproduction. It requires permission or a defense such as fair use.

So Anthropic’s right to discard books it had purchased did not, by itself, authorize permanent digital copies.

In June 2025, the Northern District of California did not answer the abstract question “May an AI read books?” It separated Anthropic’s copying into different uses.

Copies made to train particular language models. On the record before it, the court held that use to be transformative and fair.

Digital copies made from lawfully purchased print books for the internal library. The court also held the print-to-digital conversion to be fair use on the case-specific facts.

Digital copies acquired from pirate libraries and retained in a permanent, general-purpose central library. The court denied Anthropic summary judgment on fair use. That order did not finally adjudicate infringement liability on those copies; the later $1.5 billion class settlement resolved specified risks involving pirated works. It was not compensation for cutting up purchased books.

For the purchased-book scans, several facts worked together. Anthropic had paid for the paper copy. The paper copy was destroyed. One digital file replaced one physical book. The format change saved space and enabled search. And the record did not show that the converted files were sold, shared, or displayed outside the company.

That last point should not be overstated. “No evidence of external distribution” is not the same as proof of strict internal access controls. The same decision raised questions about broad engineering access to the central library, limits on internal uses, and additional copies. The narrow point is that the purchased-book PDFs were not shown to have become a public full-text service.

Destruction was not a statutory requirement. It was a fact that helped the court describe the transaction as:

One physical copy in, one digital copy out.

Now compare the alternatives.

Keep both the paper book and the scan, and the company pays for storage while two copies remain in existence. Donate or resell the paper book while retaining the scan, and the book continues to circulate—but the digital file is harder to describe as a replacement for an original still controlled by the company. License an ebook, and the company may face contractual limits, access controls, DRM, and uncertain long-term rights. Destroy the paper book, and the loose pages can move through high-speed equipment, storage costs disappear, and the company can argue that the facts resemble the one-for-one replacement accepted in Bartz.

This does not prove that destruction reduces the probability of suit or the ultimate damages bill. No empirical evidence cited here establishes that. It does show that destruction improved the defense accepted in one federal district-court ruling.

That is enough to create an incentive. Law does not need to command an act to reward it. It can change the relative cost of keeping, donating, licensing, and destroying.

The comparison cases also show the limit. Google Books and HathiTrust involved original library books that were not destroyed; full-text search, Google’s restricted snippets, and HathiTrust’s accessible formats for print-disabled users could still support fair use. In Hachette v. Internet Archive, by contrast, a one-copy-owned, one-copy-lent rule did not save a service that offered the public complete digital reading and substituted for existing book and ebook markets.

Destroying the source is therefore neither necessary nor sufficient for fair use. Purpose, access, output, and market substitution remain central. But under the facts of Bartz, the disappearance of the paper original was still legally useful. Bartz is a district-court decision on a specific record, not a nationwide safe harbor.

Digitization can preserve a text without preserving the book.

If the words and page images survive, what has been lost?

Start with the work: the author’s text, images, and expression. A good scan can usually preserve that.

Then consider the edition and the individual object. Translation, printing, paper, binding, dust jacket, foldouts, inserts, and illustrations may carry bibliographic or historical information. A particular copy may also contain a signature, inscription, marginalia, bookplate, repair, unusual binding, or provenance linking it to a person or institution.

An ISBN will not tell you any of that. A scan focused on readable pages may not capture it either.

There is also an institutional layer. A physical book can be sold, lent, donated, deposited in a public collection, or examined under future research methods. A PDF inside a private corporate library may preserve data for the company, but that is not the same as preserving public custody, independent access, or the object itself.

Scanning preserves what the capture plan was designed to capture. Keeping the original preserves the possibility that a future researcher will ask a question the scanner did not anticipate.

That does not make every ordinary copy culturally irreplaceable. A mass-produced copy with many surviving equivalents and no distinctive features may have little item-level preservation value. The question is not whether all paper must survive forever.

The question is who bears the burden before an irreversible decision.

The anger is about authority, not simply nostalgia.

Destroying books produces a stronger reaction than disposing of most corporate assets. That reaction is easy to caricature as sentimentality. It is better understood as a legitimacy problem.

First, a cultural object is being traded for private efficiency. The company sees faster scanning, lower storage costs, and a more favorable litigation position. The public sees a potentially irreversible loss converted into operating savings. Social-science research on taboo trade-offs helps explain why putting a price on something treated as culturally significant can generate moral anger even when the underlying transaction is economically rational.

Second, the gains and risks fall on different people. The model, the corpus, and the commercial benefit belong to the company. If an important edition, a copy linked to a notable person, or evidence of provenance is mistakenly destroyed, the loss is borne by authors, libraries, researchers, collectors, and future users.

The risk is also hard to calculate. Outsiders do not have the complete list, the denominator, or a way to reverse an error. Unknown, involuntary, poorly controlled, and irreversible risks tend to feel more serious than risks of the same expected monetary size.

Third, the beneficiary is also the decision-maker. If the company alone decides which copy is disposable, without a publicly reviewable standard, an independent check, or a correction process, even an error-free outcome can lack procedural legitimacy.

People are not necessarily demanding that every industrially printed book live forever. They are asking:

Who gets to decide that this copy may disappear—and what must that person prove first?

The severed spine became a powerful image because it condensed wider anxieties about AI: human knowledge accumulated over generations becomes a closed corporate asset; the public is asked to trust internal processes it cannot inspect; private benefit is concentrated while the cost of error is externalized.

Catalogs can detect risk. They cannot issue a destruction license.

“Out of print,” “rare,” and “unique” are not synonyms.

Out of print means the book is no longer ordinarily supplied by its publisher. Rarity may attach to an edition or to the features of a particular copy. “Unique” is always a limited claim: unique within the catalogs, locations, and date searched.

An ISBN identifies a publication product—a particular title, edition, and format. It does not identify the physical copy on the table. Two books with the same ISBN may be an ordinary reading copy and an author-inscribed association copy.

WorldCat is enormously useful, but it is not a census of every surviving object. Its holdings data reflects records set and maintained by participating institutions. A holding may be outdated, incomplete, inaccessible, or differently cataloged. Private and uncataloged copies may be absent.

These systems are risk signals, not proof that a copy is safe to destroy.

A workable audit at million-book scale needs three stages.

1. Freeze and document the item before cutting. Assign an item identifier. Photograph the covers, spine, title and copyright pages, printing line, edges, dust jacket, inscriptions, bookplates, marginalia, foldouts, maps, and inserts. Preserve the metadata and decision log. No destructive step should occur before this intake record exists.

2. Screen by edition, not merely by title. Cross-check national-library records, WorldCat, HathiTrust, and relevant specialist bibliographies. Flag early printing, local or minority-language publication, low institutional holdings, limited editions, unusual bindings, signatures, annotations, inserts, and conflicting metadata.

3. Route flagged items to experts. A rare-book bibliographer should assess edition and item significance; a conservator should assess material condition and scanning method. If uncertainty remains, use a cradle, planetary scanner, or camera rather than cutting the spine. The Library of Congress already treats form-feed equipment as inappropriate for fragile, high-value, fine-art, special-collection, and archival materials.

Numerical thresholds can trigger review, but they should never function as automatic permission to destroy. As a governance proposal, zero to two recorded institutional holdings should trigger a temporary halt while the holding institutions and specialists confirm the surviving copies; three to nine should require specialist review; even ten or more should not excuse an item-level check for signatures, provenance, and inserts. These are proposed risk tiers, not legal rules.

No audit can prove a universal negative. Its conclusion should be limited:

As of the stated date, within the listed catalogs, geographic scope, institutional confirmations, and physical inspection, no evidence was found that this copy was a known unique copy, a scarce edition, or an item of special significance.

Publishing a full scan may infringe copyright without authorization or a valid defense. Confidential audit access does not require public distribution of the work: an independent auditor can still review the catalog record, risk rating, pre-scan photographs, decision, sign-off, and final disposition of the physical object.

The point is not to satisfy curiosity. It is to turn an irreversible private decision into a reviewable one.

This is not a proven rare-book catastrophe. It is not a baseless moral panic either.

Public discussion has repeatedly widened the known facts.

Used and out-of-print books became “rare books.” Anthropic’s documented operation became something “all AI companies” were doing. Destruction of physical carriers became “the erasure of human knowledge.”

The evidence does not support those claims as stated.

But destructive scanning itself is not a rumor. It is a court-documented industrial process. What remains unknown is the item-level list, the physical features of the destroyed copies, the screening process, and the cultural loss, if any.

The most accurate description is a real but unquantified governance risk amplified in a low-trust environment.

That low trust predates this controversy. In a February 2026 Pew Research Center survey, 59% of US adults said they had little or no confidence in US companies to develop and use AI responsibly. This was not a survey about book destruction and cannot measure opposition to destructive scanning. It does show that the public did not encounter this story with a blank account of trust.

Musk’s statement belongs in this context. It does not prove that another company destroyed rare books. It is a market signal: a prominent AI executive publicly treated preservation of rare originals and non-destructive scanning as part of the industry’s legitimacy.

The fear may outrun some facts. It still points to a real concentration of power: a small number of companies can absorb cultural resources at enormous scale while outsiders struggle to determine what was acquired, how it was handled, what trained a model, and whether the originals still exist.

Fix the incentive, not the scanner.

The answer is not to ban digitization or preserve every common copy forever. It is to remove the rule-of-thumb advantage of proving control by destroying the source.

Start with measures that can be implemented through procurement, contracts, audits, settlements, and institutional partnerships.

Companies should be able to place originals in independent custody, seal them, or donate them to public libraries and special collections while documenting that they have surrendered commercial control of the paper copy and continue to restrict the digital file.

Current law does not establish that donation or custody is legally equivalent to destruction. This article makes the normative case that, when the disposition chain is auditable and the company cannot commercially exploit both formats, preservation should receive evidentiary treatment comparable to destruction.

Large-scale destructive scanning should also trigger item-level records, risk tiers, third-party sampling, full review of flagged materials, exception logs, and disclosure of physical disposition. A corporate corpus should preserve the chain from work to edition, item, digital file, and actual training set.

And a private PDF should not be mistaken for cultural preservation. A commercial research library needs provenance records, integrity checks, migration plans, controlled audit access, and a custody plan for shutdown, bankruptcy, or deletion. It can preserve content for a business without preserving the object or meaningful public access.

The harder reform concerns fair use itself. Section 107 does not expressly require the least harmful method. Courts could still consider whether a commercially reasonable alternative—non-destructive scanning, independent custody, or public deposit—would achieve the claimed purpose while avoiding irreversible cultural loss. That is a policy argument, not a description of current doctrine.

Copyright law did not tell Anthropic to destroy books. But under the facts accepted in Bartz, destruction helped create the cleaner legal story.

Preservation should not be the legally worse option.

Not every book must survive. But the decision about which book may disappear should not belong solely to the party that profits from its disappearance.

Primary and professional sources

Social-science background