Fair Use for Training, Infringement for Piracy, and a $1.5 Billion Settlement: What Bartz v. Anthropic Means for Copyright Owners and AI Users

In our last entry we addressed whether the output of a generative AI system can be copyrighted. This entry addresses the other side of the ledger: whether copying existing books, articles, images and code to train an AI model infringes the copyrights in those works. The first federal court to answer that question on a full record was the Northern District of California in Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.), and its answer—delivered in a June 23, 2025, summary-judgment order by Judge William Alsup—was actually two answers. Training a large language model on lawfully acquired books is fair use. However, downloading millions of pirated books from “shadow libraries” to build a permanent research collection is not. Fourteen months later, that second answer produced the largest copyright settlement in United States history: $1.5 billion, roughly $3,000 per work, paid to the authors and publishers of nearly half a million books, with final approval entered on July 21, 2026.

This post explains the decision, the settlement, and where the two now sit alongside the other AI-training cases. It then turns to what matters most to our clients: the practical effect of Bartz on copyright owners deciding whether and how to enforce, and on businesses building or buying AI systems.

For an overview of our copyright registration and licensing services, see Copyright Matters; for how we handle infringement disputes, see Litigation & Enforcement.

The Case

The Parties and the Conduct

Andrea Bartz, Charles Graeber and Kirk Wallace Johnson — a novelist and two nonfiction authors—filed a putative class action on August 19, 2024, alleging that Anthropic, the developer of the Claude family of large language models, had copied their books to train its models without permission from the authors.

Discovery established two very different sources for Anthropic’s training text. Beginning in 2021, Anthropic downloaded at least seven million books from pirate sites, including the Books3 dataset, Library Genesis (LibGen) and the Pirate Library Mirror (PiLiMi), and stored them in a central digital library from which training sets were later assembled. Separately, beginning in 2024, Anthropic purchased millions of used print books, stripped their bindings, scanned the pages, and discarded the paper—creating a digital copy of each purchased book for the same library.

Anthropic moved for summary judgment on fair use as to all of this conduct. Judge Alsup, a veteran of the Northern District’s technology docket, analyzed the record as three distinct uses rather than one, and that framing drives everything that follows.

Use One: Training the Models — Fair Use

The court held that using copyrighted books to train a large language model is fair use, and it did so in unusually emphatic terms. On the first statutory factor, the purpose and character of the use, the court found the training use “exceedingly transformative” and “spectacularly so.” The court’s analogy has been quoted in nearly every subsequent discussion of the issue:

“Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.”

The second factor (the nature of the works) favored the authors, because the books were expressive and published for a reading public. The third factor (the amount used) favored Anthropic. Despite the copying of entire books, copying the whole work was found to be reasonably necessary to the transformative purpose, and—critically—because the record showed that the models did not output the books or substantial portions of them to users. On the fourth factor, the effect on the market, the court found no cognizable harm. The authors did not contend that the models reproduced their books, and the court rejected the argument that training-generated competition—models that could help others write books—was a harm copyright law recognizes.

“The Copyright Act,” the court wrote, “seeks to advance original works of authorship, not to protect authors against competition.” The court likewise declined to credit a hypothetical market for licensing books as training data, reasoning that a copyright owner cannot defeat fair use merely by asserting a right to be paid for a transformative use.

Use Two: Digitizing Purchased Print Books — Fair Use

The court also held that scanning lawfully purchased print books into digital form was fair use. The purchase gave Anthropic the right to keep the copy. Converting it from paper to searchable, space-saving digital form, destroying the original, and using the digital copy internally rather than distributing it, was a format change that did not add a copy to the world or displace any sale. The court was careful to note that this holding depended on the print copy having been bought and then destroyed. A company that scanned a book and kept or resold the paper would face a different analysis.

Use Three: The Pirated Central Library — Not Fair Use

The third use was the one on which Anthropic lost. The court held that downloading pirated books to build a permanent, general-purpose research library—including books that were never used for training, and copies retained “should they prove useful for one thing or another”—was “its own use, and not a transformative one.” The court rejected Anthropic’s argument that later training use retroactively justified the initial piracy.

Every factor weighed against fair use for this conduct: the purpose was to acquire, for free, what could have been bought, the works were expressive, entire works were copied and retained indefinitely, and the acquisition directly displaced the sales that Anthropic itself later made when it bought print copies. The court’s language on this point was pointed. It observed that the piracy was “inherently, irredeemably infringing”, even if the copies were later put to a fair use, and that Anthropic’s own later decision to buy the books it had earlier pirated undercut any claim of necessity.

The court accordingly denied summary judgment as to the pirated library, ordered a trial on infringement and damages for December 2025, and stated that the willfulness of the conduct would be a question for the jury. Because the Copyright Act allows statutory damages of up to $150,000 per work for willful infringement, and the pirated library contained millions of works, the potential exposure was extraordinary.

Class Certification and Settlement

On July 17, 2025, the court certified a class of copyright owners of books in the LibGen and PiLiMi downloads that were registered with the Copyright Office—the first certified class in an AI-training case. Anthropic sought interlocutory review of the certification order, but on September 5, 2025, the parties announced a settlement. Its principal terms were:

• A $1.5 billion non-reversionary fund,
• A works list of approximately 482,460 eligible books, producing a gross recovery of roughly $3,000 per work—about four times the ordinary statutory-damages minimum,
• A default 50/50 split between author and publisher for each work, adjustable on proof of the parties’ contract,
• destruction of Anthropic’s copies of the pirated files and their derivatives within thirty days of final judgment, with written certification, and
• A release limited to claims arising from Anthropic’s acquisition and copying of the listed works before August 25, 2025.

Claims based on model outputs, claims concerning lawfully acquired training material, and claims for any future conduct were expressly preserved.

Judge Alsup granted preliminary approval on September 25, 2025 and, upon his taking inactive status at the end of the year, the case was reassigned to Judge Araceli Martínez-Olguín. By the May 14, 2026 fairness hearing, claims had been submitted for more than 440,000 of the eligible works—a participation rate above 90 percent, which is nearly unheard of in class practice and reflects the settlement administrator’s success in locating publishers and authors directly.

The court entered final approval on July 21, 2026. It reduced class counsel’s fee request—originally $300 million, later reduced by counsel to $187.5 million—to approximately $101.5 million, or 6.8 percent of the fund, and reduced the named plaintiffs’ service awards from $50,000 to $15,000 each. Net of fees and administration, valid claimants are expected to receive approximately $3,100 per work. Several appeals of the fee award and settlement terms were filed, and payment to class members awaits their resolution. The first distributions were expected to begin in late 2026.

One consequence of the settlement deserves emphasis. Because the case resolved before trial and before any appeal, Judge Alsup’s fair-use ruling was never reviewed by the Ninth Circuit. It remains a district-court decision—persuasive, widely cited, but not binding on any other court.

Bartz in Context: The Other Training Cases:

Bartz is best understood alongside three other decisions.

Thomson Reuters v. Ross Intelligence (D. Del. Feb. 11, 2025). Four months before Bartz, Judge Bibas, a Third Circuit judge sitting by designation, granted summary judgment against fair use where an AI legal-research start-up had used Westlaw headnotes to train a competing search tool. The court found the use not transformative because the tool served the same purpose as Westlaw and competed directly with it, and it recognized a market for licensing the headnotes as AI training data. The case did not involve generative AI, which the court itself flagged as a limitation. Ross took an interlocutory appeal. the Third Circuit heard argument on June 11, 2026, and its decision—the first appellate ruling on fair use in AI training—is expected in late 2026.

Kadrey v. Meta Platforms (N.D. Cal. June 25, 2025). Two days after Bartz, Judge Chhabria of the same court granted summary judgment to Meta on fair use for training its Llama models on books, including books obtained from the same pirate sources. But the opinion reads as a warning rather than an endorsement.

Judge Chhabria criticized Bartz for treating market competition from AI-generated works as irrelevant, articulated a theory of “market dilution” under which a model that floods the market with competing works can cause cognizable harm even without reproducing any of them, and stated that Meta prevailed only because the plaintiffs had failed to develop evidence on that theory. The opinion also stated that in most cases the fourth factor would likely favor plaintiffs who present the right evidence.

The New York Times and consolidated cases against OpenAI and Microsoft (S.D.N.Y.), together with the music-publisher case against Anthropic, Concord Music Group v. Anthropic (N.D. Cal.), and the visual-media cases against Stability AI and Midjourney, remain pending. Those cases differ from Bartz in an important respect: the plaintiffs allege that the models output their works, in whole or in recognizable part, which removes the assumption on which Judge Alsup’s third- and fourth-factor analysis rested. The Bartz settlement expressly preserved output-based claims for exactly that reason.

The Eleventh Circuit has not addressed AI training, and no Florida district court has issued a reported decision on it. The Eleventh Circuit’s fair-use jurisprudence—including its 2014 and 2018 decisions in Cambridge University Press v. Patton, which emphasized a work-by-work analysis and the importance of an available licensing market—is, if anything, more attentive to the fourth factor than Bartz, and a Florida court would not be bound to follow Judge Alsup’s treatment of the licensing-market question.

The Practical Effect of Bartz on Copyright Law:

Stripped of the headlines, Bartz has changed copyright practice in six identifiable ways.

1. Acquisition is now the battleground, not training

The most durable lesson of Bartz is that how a work was obtained can be analyzed separately from what was done with it. Two federal judges in the same courthouse, within two days, treated training on books as fair use, and the Third Circuit is the only appellate court close to ruling. Copyright owners contemplating a training-based claim should assume that a transformative-use finding is likely on the training itself, and should focus instead on the acquisition: Was the work downloaded from a pirate site? Was a paywall, terms-of-service restriction or technical measure circumvented? Was a licensed copy used beyond the license? Was a copy retained after the training purpose ended? Each of those facts supports a claim that survives a fair-use finding on training, and each is a fact within the defendant’s records rather than the plaintiff’s speculation. Conversely, developers should assume that a lawful-acquisition record is the price of a fair-use defense.

2. Statutory damages give registered owners real leverage

The $1.5 billion figure is a product of arithmetic: statutory damages of $750 to $150,000 per registered work, multiplied by hundreds of thousands of works, in front of a jury that would hear the word “pirated.” That leverage exists only for works registered with the Copyright Office before the infringement began or within three months of publication, as 17 U.S.C. § 412 requires, and the Bartz class was limited to registered works for precisely that reason.

Authors, publishers, photographers, software developers and other owners who have not registered their catalogs have, in practical terms, opted out of the most valuable remedy the Copyright Act provides.

We discuss registration strategy on our Copyright Matters page; for owners with large back catalogs, group registration options can make comprehensive registration affordable.

3. A licensing market for training data now exists — and it cuts both ways

Judge Alsup declined to recognize a market for licensing books as training data. But the market has since materialized: AI developers have entered paid agreements with news organizations, publishers, stock-image libraries, music companies and user-generated-content platforms, and the Bartz settlement itself put a price on unlicensed acquisition.

The existence of that market strengthens the fourth-factor argument for future plaintiffs (an argument Judge Chhabria’s Kadrey opinion invites), and it gives owners a commercial alternative to litigation. Owners with content that developers want—clean, well-organized, rights-cleared collections—are increasingly able to license it on terms that include attribution, output restrictions and audit rights, none of which a court can award.

4. Output claims are the next frontier, and Bartz did not resolve them

Every fair-use holding in Bartz rested on the finding that the models did not reproduce the plaintiffs’ books. Where a model does reproduce protected expression—memorized passages, near-verbatim summaries, images that replicate a photographer’s composition, code that reproduces a licensed library—the analysis is different, and the settlement expressly preserved those claims. Owners who can show that a model outputs their work should plead that as a separate infringement, distinct from training, and should preserve the evidence (prompts, outputs, dates and model versions) at the time it is discovered.

5. Format-shifting of purchased copies is on firmer ground

The print-to-digital holding has implications beyond AI. It supports the position that an entity that lawfully owns a physical copy may convert it to digital form for its own internal use, provided the original is not retained or redistributed and the digital copy is not shared. Libraries, archives, law firms and businesses with large paper collections should note the conditions the court attached: one copy in, one copy out, and no distribution.

6. The class action has arrived in copyright

Copyright class actions were historically rare because ownership, registration and damages vary work by work. Bartz showed that where the defendant’s own records identify the works copied and the source of each copy, those obstacles can be overcome, and a settlement administrator can locate the owners of nearly half a million books. Owners of individual works now have a realistic path to recovery without individual suits, and developers face aggregated exposure that individual settlements cannot cap.

Recommendations

For Copyright Owners

• Register the catalog. Timely registration is the prerequisite to statutory damages and attorney’s fees, and it was the ticket into the Bartz class. Prioritize works that are commercially significant or likely to appear in training corpora, and use group registration for serials, photographs and short works.
• Determine whether your works were in the pirate datasets. The Books3, LibGen and PiLiMi datasets have been analyzed publicly, and the Bartz works list is searchable through the settlement administrator. Owners who did not submit a claim in Bartz but whose works were acquired from the same sources by other developers retain claims against those developers, and the release in Bartz does not affect them.
• Investigate acquisition, not just training. When a claim is contemplated, discovery should target the source of the copy, the date of acquisition, the retention period, and any circumvention of access controls. These are the facts that defeated fair use in Bartz and that no training-purpose argument can cure.
• Build the fourth-factor record. If the theory of harm is market substitution or dilution, retain an economist early and develop evidence of actual licensing transactions, lost sales and output-based competition. The Kadrey plaintiffs lost for want of that evidence, and the Bartz court found none.
• Consider licensing before suing. A negotiated license can deliver terms—attribution, output limits, audit rights, future-version coverage—that litigation cannot, and the existence of a licensing relationship strengthens the market-harm argument against those who do not license.
• Preserve output evidence. Screenshots, prompts, dates and model identifiers documenting reproduction of your work are the foundation of an output claim, and models change frequently.

For Businesses Building, Fine-uning or Deploying AI:

• Source training data lawfully, and document it. Purchase, license or use openly licensed material, record the source, date and license for every dataset, and do not ingest content from shadow libraries or through circumvention of paywalls or terms of service. The distinction between Bartz’s two outcomes is a distinction in provenance.
• Do not build a general-purpose library. The court treated retention of pirated works for possible future use as a separate infringing act. Acquire what is needed for a defined purpose, and dispose of or segregate what is not used.
• Test for and prevent regurgitation. The fair-use holdings depended on the finding that the models did not output the works. Implement and document output filters and memorization testing, because an output claim is the one the Bartz settlement preserved.
• Diligence vendors and models. Companies deploying third-party models inherit the provenance risk of their vendors. Contracts with AI providers should include representations regarding training-data sourcing, indemnities for infringement claims, and cooperation obligations in litigation. Our technology and start-up and entertainment law practices routinely negotiate these terms.
• Watch the Third Circuit. The Thomson Reuters v. Ross decision expected in late 2026 will be the first appellate word on fair use in AI training, and it will be decided by a court that has recently emphasized market harm. Businesses should not treat Bartz as settled law.

Conclusion

Bartz v. Anthropic is at once the most developer-friendly and the most owner-friendly decision yet issued on AI training. It holds that training a model on lawfully acquired books is a transformative fair use, language that will be quoted by every AI defendant for years. But it also holds that obtaining those books by piracy is an ordinary, willful infringement for which the Copyright Act’s statutory damages apply in full—a holding that, multiplied across half a million works, produced a $1.5 billion settlement and a court-ordered destruction of the pirated files.

The practical effect is to shift the fight from what AI does with copyrighted works to how it obtains them, to make timely registration the difference between a claim worth thousands of dollars per work and a claim worth nothing, and to leave for another day—and, likely, for the Third Circuit and the output cases—the questions Bartz did not reach.

If your works may have been used to train an AI system, or if your business is building or deploying one, our attorneys can assess the exposure and the opportunities under Bartz and the cases that have followed it. Please visit our Copyright Matters and Litigation & Enforcement pages, or contact us at 305-374-8303.