Federal judge rules it's 'fair use' for Anthropic to destroy rare books to feed to AI language models

The ruling does change the fundamental nature of what the books were intended for: they were intended to be read by human beings, not used as training for machines to replace human thought and ingenuity.

AI companies are buying old books, scanning them, then destroying them. Anthropic has been buying the books to feed the machine with human knowledge and ideas written down before the AI era, so only books with publication dates before 2022 will do. Once they have the books, they rip off the spines, feed them through a scanner, and then on into the gaping maw of the AI LLMs that consume everything, including, eventually, themselves.

The revelations came out in a lawsuit brought by authors who said the destruction violated the Copyright Act. Federal judge for the Northern District of California William Alsup ruled that the bulk buy, scanning, and destruction is fair use under Section 107 of the Copyright Act. As far back as 2024, Anthropic, which has come under fire by the Trump administration over national security concerns, said "We don’t want it to be known that we are working on this." It's not a good look to be destroying books.

The suit reads that authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson were a few of the authors whose books were bought by Anthropic. For use in Claude, Anthropic "assembled these copies into a central library of its own, copied further various sets and subsets of those library copies to include in various 'data mixes,' and used these mixes to train various LLMs. Anthropic kept the library copies in place as a permanent, general-purpose resource even after deciding it would not use certain copies to train LLMs or would never use them again to do so." The authors had not agreed to this and these copies of their books were taken permanently out of circulation, never to be read by human eyes again.

Anthropic is one of these companies that has been buying physical books in bulk to rip apart, scan, then destroy, in service to its AI model Claude. Any books will do, even if that book is the very last copy of that book in the world. Legally, this is allowed under the first-sale doctrine, which permits a book-buyer to do whatever they want with that book object, without a need for the copyright owner's permission. It's an object, like a pair of shoes or a blender.

The ruling reads: "every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy. The print original was destroyed. One replaced the other. And, there is no evidence that the new, digital copy was shown, shared, or sold outside the company."

In the suit, a judge found that because Anthropic was using the books in a way that fundamentally transformed them into something else, in this case turning them from books into raw data that will have no reference to their original self, and that this constituted fair use. Previous to the bulk buying of books, Anthropic pirated books or bought pirated books. They looked to former head of Google Books' scan project Tom Turvey to come in and get more data to feed the LLM. That data was not just information but also style and tone, proprietary authorial elements.

There's lots of companies that are willing to do the selling. ISBNdb claims to have the "world's largest book database," and says "the world's best AI training data is sitting on a shelf." They also promise not to reveal who is doing the bulk buying, knowing that no AI company wants to be the subject of a headline about how they're destroying millions of books at a go.

ISBNdb says that AI training on AI material results in data degradation. "Not all data degradation is intentional," says ISBNdb. "When AI systems train on text that was itself AI-generated, a documented phenomenon called model collapse occurs: subtle linguistic nuances vanish, systematic errors compound, and outputs converge on repetitive patterns. Each generation trained on synthetic data is slightly worse than the last. Print books from the pre-LLM era are structurally guaranteed to be free of this contamination. That alone is a significant advantage."

404 Media spoke to a bookseller who said that bulk orders from his store have resulted in the destruction of out-of-print books for which there are only single copies remaining. "I personally have mixed feelings about all of this," said the bookseller. "It benefits me financially as well as by clearing out old inventory that is otherwise unlikely to sell. I’ve been well-suited for these sales with inventory from overseas and foreign language books. On the other hand, I don’t like the end-use, and I don’t like that uncommon books are being pulped."

The ruling reads: "This order grants summary judgment for Anthropic that the training use was a fair use. And, it grants that the print-to-digital format change was a fair use for a different reason. But it denies summary judgment for Anthropic that the pirated library copies must be treated as training copies." The ruling does change the fundamental nature, however, of what the books were intended for: they were intended to be read by human beings, not used as training for machines to replace human thought and ingenuity.

As it stands, Anthropic and other AI companies that are developing and furthering their large language models for use both as information and writing tools are permitted to take books out of circulation, feed them to the LLM, and destroy them, without concern for copyright law or, as it turns out, the preservation of the scope and breadth of human history, knowledge, experience, and creativity.

I just had an extensive research discussion with AI on this article and the larger issue it reports. Below is an AI generated summary of my research discussion:

The Uneasy Bargain: Anthropic, Fair Use, and the Constitutional Promise to Authors
Introduction
In June 2025, Senior Judge William Alsup of the Northern District of California handed down a ruling that will likely be remembered as one of the founding legal documents of the artificial intelligence era. In Bartz v. Anthropic, Alsup held that Anthropic's practice of buying used print books, scanning them into digital form, and destroying the physical originals was "exceedingly transformative" fair use under the Copyright Act — even as he found the company's separate practice of downloading millions of pirated books to be unlawful. The ruling was narrow in its holding but sweeping in its implications, and it reopened a two-hundred-year-old argument buried in eighteen words of the Constitution: what did the Framers actually mean when they gave Congress the power "to promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries"?
This essay traces that argument as it unfolded — from the ruling's narrow legal mechanics, through the doctrinal history of fair use, into the deeper constitutional and historical dispute over what the Copyright Clause was actually built to protect, and finally to the practical, measurable stakes now emerging as AI systems begin to displace the very authors whose work trained them.
The Ruling Itself: A Split Verdict
Alsup's opinion did not bless everything Anthropic had done. He drew a sharp distinction between two acts. The first was digitization: Anthropic spent millions of dollars purchasing used physical books, stripped their bindings, scanned their pages, and discarded the originals, retaining only the internal digital files. Alsup likened this to "a reader aspiring to be a writer" and, more concretely, to ripping a purchased CD onto an iPod — the digital copy simply replaced the print copy the company already owned outright under the first-sale doctrine, without redistribution to any third party. On this narrow act, all four fair-use factors, in his view, favored Anthropic: no market substitute was created, nothing was shared externally, and the company could legally have kept the print books indefinitely without ever digitizing them.
The second act was different. Anthropic had also downloaded more than seven million pirated books to build a permanent internal library, and Alsup found this was not excused by any downstream training use — "creating a permanent, general-purpose library was not itself a fair use excusing Anthropic's piracy," he wrote. That claim proceeded toward a damages trial and, ultimately, a settlement.
The public reporting on the case, including a widely circulated Post Millennial headline describing the destruction of "rare books," overstated and mischaracterized this holding in a subtle but important way. The books at issue were not rare or antiquarian — they were ordinary used paperbacks purchased in bulk — and the headline collapsed two distinct rulings (fair use for purchased books, infringement for pirated ones) into a single, misleadingly broad claim that the judge had blessed Anthropic's book-sourcing practices wholesale.
Fair Use as Doctrine: Broader and Narrower Than It Appears
A natural first reaction to the ruling is to ask how "fair use" — an exception meant to permit criticism, commentary, and scholarship — could possibly cover the wholesale digitization of an entire copyrighted work. But fair use has never been confined to excerpts. The Supreme Court's 1984 decision in Sony Corp. v. Universal City Studios upheld home videotaping of entire television broadcasts as fair use, reasoning that copying a whole work does not automatically defeat a fair-use defense when market harm is minimal. Three decades later, the Second Circuit's decision in Authors Guild v. Google, left standing by the Supreme Court, held that Google's scanning of more than twenty million entire books for its search index was fair use because the tool augmented public knowledge without providing "a substantial substitute" for the books themselves.
Alsup's ruling extends this same transformative-use logic to AI training, and in doing so it tracks a fifteen-year-old appellate precedent rather than inventing a novel doctrine from nothing. But the comparison to Authors Guild v. Google also exposes the weakest point in Alsup's reasoning. Google Books never generated new prose that competed with an author's voice; it returned brief, disjointed snippets. A trained language model can produce long-form text, answer detailed questions about a book's arguments, and in some documented instances reproduce near-verbatim passages — a categorically more substitutive capability than a card-catalog search tool. Whether "training" is more like Google's narrow indexing function or more like an unlicensed engine for reproducing an author's expression is the unresolved question beneath the settled-sounding legal label.
The commercial-use objection deserves equal scrutiny. It is tempting to think that because Anthropic profits from Claude, any unlicensed use of copyrighted material to build it should automatically fail the fair-use test. But the Supreme Court explicitly rejected that reasoning in Campbell v. Acuff-Rose Music (1994), holding that commercial purpose is not presumptively unfair — it is weighed against transformativeness and market harm, not treated as an automatic disqualifier. The clearest illustration of where commercial motive did sink a fair-use defense is A&M Records v. Napster (2001), where the Ninth Circuit found that even non-monetized peer-to-peer file sharing constituted commercial exploitation because it substituted directly for the purchase of a CD. The distinction the courts have consistently drawn is not "did money change hands" but "did the copy displace the market the copyright holder was entitled to exploit." Napster's users took "the whole and the heart" of a song and gave nothing back but a costless substitute. Anthropic's book-training use, Alsup concluded, does not.
Ownership, Not License: The First-Sale Doctrine
A recurring intuition in objections to the ruling is that buying a book is merely licensing the right to read it, so a purchaser has no further right to reproduce, alter, or destroy the physical object. This intuition, though understandable, inverts how American copyright law actually treats the sale of a physical copy. The first-sale doctrine, codified at 17 U.S.C. § 109, provides that once a copyright holder sells a physical copy, the buyer owns that particular copy outright and may resell it, lend it, or destroy it without the copyright holder's permission. Software and ebooks are typically licensed rather than sold, which is why first-sale rights usually do not attach to them — but Anthropic was buying used print books outright. Its legal right to destroy those books after digitizing them required no fair-use justification at all, because destroying one's own lawfully purchased property is not, standing alone, an act copyright law regulates. Publishers' copyright-page disclaimers — the familiar "no part of this publication may be reproduced" boilerplate — carry no independent legal force beyond providing notice that a work is copyrighted; they cannot contractually override a federal statutory right such as fair use or first sale.
The intermediate-copying question — that Anthropic's digitization began as an exact, unmodified copy of a purchased book — might seem to categorically foreclose a fair-use defense on its face. But Sega v. Accolade (1992) establishes that an exact, wholesale intermediate copy does not defeat fair use if the ultimate purpose and effect of the copying is transformative and non-substitutive. Accolade copied Sega's object code byte-for-byte as the first step in reverse-engineering game compatibility, and the Ninth Circuit still found fair use because Accolade's finished product did not contain Sega's code and did not compete in Sega's market. Courts look through the mechanical fidelity of the copying step to what it was for and what effect it produced — though critics note that Sega involved copying to extract unprotectable functional elements, whereas books are pure expressive content sitting at the historical core of what copyright protects, a distinction Alsup himself acknowledged weighed against Anthropic on the "nature of the work" factor even as he ruled for the company overall.
The Constitutional Question: What Did "Securing... Exclusive Right" Actually Mean?
Beneath the statutory doctrine lies the harder constitutional question of what the Framers meant to secure. The Copyright Clause does not grant authors an unconditional property right; its own text frames the exclusive right as the means Congress may use toward the end of promoting "the Progress of Science and useful Arts." The Supreme Court's 2003 decision in Eldred v. Ashcroft treated fair use as one of the "traditional contours" of copyright that make the resulting speech restriction constitutionally tolerable in the first place — not an infringement carved out despite the Constitution, but a mechanism built into copyright's design from the start.
Yet this reading, however doctrinally settled, sits uneasily against a serious objection: American copyright cannot simply be read as an extension of the English legal tradition it borrowed from, because that tradition was itself a product of a very different constitutional order. English printing control before 1710 had nothing to do with authors at all. In 1557, the Stationers' Company received a Royal Charter from Mary I granting the guild a printing monopoly explicitly designed to suppress seditious and heretical material — a tool of Crown censorship, not creator protection. Even the 1710 Statute of Anne, the direct ancestor of the American clause, emerged only because the Stationers, facing the loss of their monopoly when Parliament refused to renew licensing controls in 1694, strategically pivoted to argue that authors — rather than the guild — should hold the underlying right, a tactical move that happened to align with, but was not originally motivated by, respect for individual creative labor.
An American Constitutional Convention founded on a revolution against monarchical property control had every reason to complete, rather than merely import, that shift toward genuine individual ownership — to treat the exclusive right not as a residual grant permitted by a sovereign but as a real, personally held entitlement, secured against confiscation by the state or by the public, robust enough that a creator could reasonably expect to profit from years of labor before any of it passed into public use. This reading finds direct support in Madison's own words, the only Framer's contemporaneous account of the clause's purpose. In Federalist No. 43, Madison wrote that "the public good fully coincides… with the claims of individuals" — not that public benefit justified diminishing the author's claim, but that protecting the claim and serving the public were the same act. "The utility of this power," he wrote, "will scarcely be questioned."
Madison's brevity on this point is often mistaken for a lack of substance, but read against the clause's own text, it resolves what looks at first like a tension between two goals — private property protection and public scientific advancement — into a single, unified design. The clause does not present these as competing interests to be balanced against one another; it presents securing the exclusive right as the mechanism by which progress is promoted. The two goals are not separate ends the Framers hoped would coincidentally align. They are cause and effect: promotion of progress is the outcome the exclusive right is built to produce, and the exclusive right is the tool by which the Framers expected authors and inventors to keep taking the risk of creating in the first place.
This is where the stifling-versus-promoting distinction becomes decisive to reading the clause correctly. A regime that treats the "Progress" language as an independent license for the public to appropriate or share a creator's work — on the theory that spreading knowledge itself promotes science — inverts the clause's actual logic and risks producing the opposite of its intended effect. If creators anticipate that their work will be taken from them and redistributed for others' benefit as soon as it proves useful, the rational response is to create less, not more, or to conceal work rather than publish it. Progress is stifled, not promoted, whenever the expectation of confiscation outweighs the expectation of reward. The exclusive right exists precisely to foreclose that outcome: by guaranteeing the creator a real and defensible opportunity to profit for a defined period, the clause removes the disincentive that would otherwise suppress investment in the "Science and useful Arts," and only after that incentive has done its work does the knowledge pass, secure and undiminished, into the public domain. Read this way, "promoting Progress" and "securing the exclusive Right" are not two goals in tension that Madison happened to notice coincided — they are one goal, described in cause-and-effect terms, which is exactly why Madison could say the public good and the individual's claim "fully coincide" without treating that as a remarkable or debatable proposition. The public good was never separate from the author's incentive to begin with; it was the anticipated consequence of protecting that incentive faithfully.
The clause's core purpose comes into clearer focus under this reading: to prevent the stifling of creative and scientific investment by guaranteeing creators a genuine, protected opportunity to profit from their work — not to mandate that their creations be shared or redistributed the moment they are produced. The incentive is the mechanism; the public benefit is the intended byproduct of a well-functioning incentive, not a separate entitlement that the public may seize directly from the author's labor.
The Unresolved Question: Where Does Training Fall?
None of this settles whether AI training on copyrighted books is the kind of "narrow, non-substitutive exception" that even an incentive-centered reading of the clause could tolerate, or whether it is closer to the wholesale appropriation the clause was designed to prevent. The 1787 record does not answer this directly, because nothing resembling fair use as a distinct doctrine existed at the Founding — it developed through case law only in 1841, in Folsom v. Marsh, and Congress did not codify it until 1976.
What has changed is the evidence available to answer the practical version of the question. Alsup's 2025 ruling rested on a record showing no direct market substitution: Anthropic paid for every book it retained, and nothing was redistributed. But data emerging through 2026 complicates that premise in ways the original ruling could not have anticipated. Field research shows AI-generated answer summaries reduce outbound clicks to original sources by nearly 40 percent, and industry-wide data shows publisher referral traffic falling by roughly a third within a single year as AI answer engines increasingly satisfy a reader's question without ever directing them to the underlying work. Reuters Institute data shows only 4 percent of AI chatbot users click through to a cited source at all — meaning the citation mechanism many AI systems provide functions, in practice, as an unused formality rather than a live channel of credit or compensation back to the creator.
This is significant because market effect is, in the Supreme Court's own words from Harper & Row v. Nation Enterprises (1985), "the single most important element of fair use." If future litigation can show — as the 2026 traffic data increasingly suggests at an industry level — that AI-generated answers measurably displace engagement with, and compensation for, specific authored works, that evidence would speak directly to the factor Alsup found lacking in Anthropic's 2025 record. The doctrinal label "transformative" may accurately describe the training act examined in isolation, but the deployed product's real-world effect on the market for original authorship is an empirical question the summary-judgment record did not, and perhaps could not yet, fully capture.
Conclusion
The dispute over Bartz v. Anthropic is not really a disagreement about whether the judge read a single word of the Constitution correctly. It is a disagreement about which of two coherent constitutional visions should control: one that treats the exclusive right as the load-bearing guarantee the Framers built to make creative and scientific labor worth the risk, and one that treats transformative public use as an equally load-bearing counterweight built into copyright from its English origins onward. Both visions can claim genuine roots in the historical record. What is emerging now, as AI systems move from training to deployment at civilizational scale, is the practical evidence that will eventually force courts, and likely Congress, to decide which vision the Constitution's "limited Times" and "exclusive Right" were actually built to protect — and whether the incentive Madison assumed "scarcely need be questioned" can survive a technology the Framers had no reason to imagine.

-PJ

p.s.

Cogley: Books, young man, books. Thousands of them. If time wasn't so important, I'd show you something. My library. Thousands of books.

Captain James T. Kirk: And what would be the point?

Cogley: This is where the law is, not in that homogenized, pasteurized, synthesized... . Do you want to know the law, the ancient concepts in their own language, learn the intent of the men who wrote them, from Moses to the tribunal of Alpha 3? Books.

Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.