Foreword

On July 20, 2026, Judge Araceli Martínez-Olguín of the U.S. District Court for the Northern District of California officially approved the $1.5 billion settlement agreement reached between Anthropic and a collective of authors in the case Bartz v. Anthropic. This is one of the largest copyright class-action settlements in U.S. history, and it represents the clearest judicial clarification to date for the generative AI industry on the question of “where training data comes from.”

For developers and AI practitioners, the value of this case does not lie in the clickbait narrative of “AI companies losing another lawsuit,” but rather in the court splitting two previously intertwined legal issues apart: Training large models with copyrighted works and bulk downloading these works from pirate sites are not legally the same thing. The former was ruled as Fair Use by Judge William Alsup in June 2025; the latter constitutes infringement, which ultimately led to this record-breaking settlement.

This article sorts out the case background, settlement terms, and actual industry impacts based on public sources including court orders, AP, Ars Technica, Authors Guild, and others.

Case Background: Where Did Claude’s Training Data Come From?

Bartz v. Anthropic was filed in 2024 by authors including thriller novelist Andrea Bartz and Kirk Wallace Johnson, alleging that Anthropic used their copyrighted works without authorization while developing the Claude large language model. The case quickly became a landmark in AI copyright litigation—it challenged the training inputs, not whether the model’s outputs copied original works; the plaintiffs did not claim that Claude produced infringing content when answering user questions.

According to Judge William Alsup’s June 23, 2025 Fair Use ruling (787 F. Supp. 3d 1007), Anthropic’s methods of obtaining training data can be roughly divided into three categories:
1. Bulk downloading from pirate sites. In June 2021, Anthropic employee Mann downloaded at least 5 million pirated e-books from Library Genesis (LibGen); in July 2022, he downloaded at least 2 million more from Pirate Library Mirror (PiLiMi). Anthropic was aware that these sources were pirate sites. The total number of pirated downloads exceeded 7 million.
2. Scanning legally purchased physical books. Anthropic purchased large quantities of second-hand paper books, unsealed them, scanned them, and digitized them into its internal “central library”.
3. Using known datasets such as Books3, which contains approximately 183,000 works.

Anthropic’s goal, in Judge Alsup’s own words, was to build a central library of “all the world’s books” and permanently retain it. Engineers then selected different subsets from this repository, cleaned and tokenized them for training each version of its LLM.

2025 Critical Ruling: Training is Fair Use, Pirated Downloads Are Not

Before approving the settlement, the court issued a split ruling on the core legal issues, which is key to understanding the $1.5 billion settlement amount.

AI Training Itself: Fair Use

Judge Alsup determined that Anthropic’s use of copyrighted books to train Claude and its predecessor models constituted “highly transformative” use, meeting the four-factor Fair Use analysis under Section 107 of U.S. Copyright Law, constituting fair use. The court specifically noted that training LLMs does not mean providing users with the original book content, but rather learning linguistic patterns and statistical regularities, whose purpose is completely different from the original published works.

Aparna Sridhar, Deputy General Counsel of Anthropic, still emphasized this ruling after the settlement was approved: “Training AI using books constitutes fair use under copyright law.” For AI companies, this is one of the most favorable judicial statements to date.

Pirated Downloads and the Central Library: Infringement

However, the court reached the opposite conclusion regarding downloading and permanently preserving pirated copies from LibGen and PiLiMi. Judge Alsup wrote that anyone who copies copyrighted material from pirate sites “has already infringed, full stop.” Even if these copies might later be used for LLM training—and training itself was ruled as Fair Use—the act of acquiring and retaining pirated copies still cannot be exempted.

The court further distinguished the scenario of scanning after legally purchasing physical books: Replacing purchased paper books with digital copies, for internal retrieval only, without increasing the number of copies or distributing them externally, constitutes Fair Use. But “first pirate, then pay for the same work” does not offset the infringement: “Paying later for the same work cannot undo the harm caused by the pirated copy.”

Therefore, when the case entered the damages phase, Anthropic won on “whether training was legal” but lost on “how the data was obtained”—the settlement targeted the latter.

$1.5 Billion Settlement: Terms and Figures

In September 2025, Judge Alsup preliminarily approved the settlement plan; he later retired, and Judge Martínez-Olguín took over and completed the final approval. The main terms of the settlement are as follows:
1. Settlement Fund: Anthropic contributes \(1.5 billion (a nonreversionary settlement fund) for eligible copyright holders to claim. 2. **Compensation per Work**: Approximately **\)3,000 per work (some reports specify $3,351.39), about four times the statutory minimum damages. Judge Martínez-Olguín considered this ratio “fair.”
3. Coverage Scope: Approximately 506,000 works are included in the class action; as of the approval date, more than 91% of affected authors and publishers have submitted claims, involving more than 440,000 books.
4. Destruction of Pirated Copies: Anthropic must destroy pirated book copies obtained from LibGen, PiLiMi and other sources; authors may file separate lawsuits if the works are misused again after the settlement.
5. Opt-Out of Class**: 350 class members successfully opted out and chose to file lawsuits individually; the court rejected most late opt-out requests (the opt-out deadline was March 30, 2025).

Attorney Fees and Lead Plaintiff Awards Reduced

The judge approved the \(1.5 billion total amount, but cut portions of the allocation: - Attorney Fees: Class Counsel originally requested **\)187.5 million (12.5% of the settlement fund), which the court reduced to approximately $101.6 million (about 6.8%).
- Service Awards for the three lead plaintiffs: Reduced from the requested \(50,000 per person to **\)15,000 per person, as the judge deemed higher amounts “unreasonable.”

If there is a remaining balance in the fund after all valid claims are paid, the parties expect to redistribute the surplus to class members unless it is economically unfeasible.

Why It’s “Largest in History”: Comparison with Other Industry Lawsuits

Both the Authors Guild and AP stated that this is the largest known copyright class-action settlement. The Bartz case is also the first major case among dozens of AI copyright lawsuits to reach a significant settlement—companies including OpenAI, Meta, and Stability AI still face similar lawsuits and have not reached comparable final resolutions.

For Anthropic, the settlement means:
- Locking in the cap: Avoiding the uncertainty of damages trials (statutory damages can reach up to $150,000 per work, and the theoretical exposure for 7 million pirated copies is staggering).
- Preserving the Fair Use victory: The ruling on the legality of training is not affected by the settlement and can still be cited as precedent.
- Concretizing compliance costs: $1.5 billion marks the price of “taking shortcuts”—the industry previously viewed the cost of pirated training data mostly as reputational risk, but now it is a calculable financial risk.

The lead plaintiffs said in a statement via Reuters: “This brings us closer to true accountability for Anthropic and sends a warning to all AI companies: they cannot take legal shortcuts or override creators’ rights.”

Three Insights for AI Training Data Compliance

Combined with the Bartz case ruling and settlement, AI companies and data engineering teams should at least pay attention to the following divisions of labor:

Fair Use protects the method of use (transformative training), and does not automatically whitewash acquisition methods (piracy). Bulk scraping of copyrighted content from LibGen, PiLiMi, BT sites, or unauthorized crawlers, even for internal training only, may still constitute independent infringement. Compliance review should be split into two gates: whether the data source was legally obtained; whether the use after acquisition complies with Fair Use—both gates must be passed.

2. Internal “Permanent Libraries” Amplify Risks

The core act that led to Anthropic’s infringement was not only “downloading pirated books” but also long-term retention as a general research resource. The court made it clear that even if Anthropic later decided not to use these pirated copies for LLM training, the retention act itself was still unreasonable. If data teams build a central corpus, they need to have retention strategies, source audits, and deletion mechanisms to avoid the model of “download once, hold forever”.

Alsup’s ruling endorsed Fair Use for “purchasing physical books → scanning → replacing storage”, but large-scale book scanning still has practical and contractual restrictions (the boundary of the first sale doctrine, prohibited scanning clauses in publishing contracts, etc.). More secure paths include: signing training licenses with publishers or collective management organizations, using authorized datasets (such as some academic corpora, publisher partnerships), or synthesizing/self-building data. The Bartz settlement shows that the licensing fees saved may come back as damages multiples higher in magnitude.

Conclusion: Precedent Set, Industry Unfinished

Judge Martínez-Olguín emphasized in the approval order that about 95% of class members received notice, and the high participation rate indicates that most authors considered the settlement “fair.” Both Anthropic and the authors’ lawyers hope to distribute the fund and conclude the lawsuit as soon as possible.

But for the entire generative AI ecosystem, the story is far from over. The Fair Use training doctrine is currently mainly supported in the Anthropic case in the N.D. Cal., and other courts and other causes of action (such as output-side infringement, DMCA, state law) are still pending. Regulatory frameworks such as the EU AI Act and national text and data mining (TDM) exceptions are also shaping compliance boundaries in parallel.

For Claude users and ordinary developers, the short-term impact is limited: Claude products will not be taken off the shelves, and the training Fair Use ruling remains valid. For AI company data leaders, this is a clear pricing signal—the $1.5 billion bought the settlement for pirated data, not a denial of the legality of training itself. Next, the focus of industry competition will shift from “who can download more books” to “who can build a sufficiently large corpus in a compliant way”, and whether the Fair Use precedent will hold up in appeals and sister cases.

References