This Is the Largest Copyright vs. Ai Settlement in History: How Anthropic Paid 1.5 Billion Dollars for Pirated Books
In September 2025, a precedent occurred that permanently changed the relationship between artificial intelligence and copyright. AI company Anthropic – creator of the advanced language model Claude – reached a settlement of 1.5 billion dollars plus interest. This is the largest publicly disclosed copyright settlement in the history of the United States.
As lawyers, we must be very clear: these are not just impressive figures. This is a message to the entire AI industry that the era of the “I will ask for forgiveness instead of permission” approach is over.
How It All Started – The Anatomy of a Pirated Download
In August 2024, three authors – Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber – filed a class‑action lawsuit against Anthropic before the District Court for the Northern District of California. They claimed that Anthropic had massively downloaded books from pirate websites to train its AI model, Claude.
Specifically, the lawsuit alleged that Anthropic:
- downloaded at least about five million books from the Library Genesis (LibGen) website,
- downloaded at least about two million books from Pirate Library Mirror (PiLiMi),
- In total downloaded more than seven million pirated copies of books.
These books were not used once and then forgotten. Anthropic organised them and kept them in an internal “central library” – a shadow library that served as a source for model training. It is particularly important that the company was aware that these were pirated copies, yet chose to use them instead of paying for licenses or buying legitimate digital copies.
The Pivotal Decision of Judge Alsup
Before the settlement, Judge William Alsup issued a decision in June 2025 that laid the legal foundations for the entire case. His decision was a so-called “split decision” – in one part, he sided with the AI industry, and in another part he drew a firm red line.
The judge drew a clear distinction between:
- The use of lawfully acquired books to train AI models,
- downloading and storing pirated books from sites such as LibGen and PiLiMi.
For lawfully acquired books (purchased or properly licensed), the judge found that their use for model training constitutes “quintessentially transformative use” and falls under the fair use doctrine. In that part, he accepted the AI industry’s argument that model training is a new, transformative purpose that differs substantially from the original reading or distribution of the work.
However, when it comes to pirated sources, Judge Alsup was completely unambiguous: downloading millions of books from websites known to distribute pirated content constitutes copyright infringement. The court found that none of the fair use factors weighed in favour of Anthropic when pirated content was at issue:
- the purpose and character of the use were compromised by the manner in which the content was obtained,
- obtaining from illegal sources cannot be justified by a transformative purpose,
- the scale (millions of copies) indicates systematic, deliberate infringement,
- the market effect is clearly negative because pirated copies substitute for legally paid copies.
The judge drew a crucial distinction: it is not enough for the use to be transformative; the manner of acquiring the material must also be lawful. In other words, the end (developing advanced AI) does not justify the means (piracy).
Theoretical Liability of 75 Billion Dollars
To understand why the settlement reached 1.5 billion dollars, we need to look at the U.S. regime of statutory damages in copyright law.
Under U.S. copyright law:
- the basic range for each work is from 750 to 30,000 dollars per work,
- for willful infringement, the amount can go up to 150,000 dollars per work,
- for “innocent” infringement, it can be reduced to 200 dollars per work.
In this case, the class action covered about 500,000 works. If we applied the maximum amount for willful infringement (150,000 dollars per work), Anthropic’s theoretical maximum liability would be around 75 billion dollars.
In that light, the settlement of 1.5 billion dollars begins to look like a compromise that was a costly blow for the company, but still far more favourable than the risk of a full judgment.
Settlement Terms: Money, Data Destruction, and Scope
The settlement in Bartz v. Anthropic includes both financial and non‑financial obligations.
Financial component:
- Anthropic pays 1.5 billion dollars into a settlement fund, with interest accruing,
- rights holders are roughly compensated with about 3,000 dollars per work for the first 500,000 works,
- if the number of qualifying works proves higher, additional per-work payments are envisaged,
- class counsel receives about 25% of the settlement fund in attorneys’ fees (approximately 375 million dollars).
Non-financial obligations:
- Anthropic must promptly and permanently delete all pirated files downloaded from LibGen and PiLiMi, as well as their internal copies,
- the settlement applies only to past conduct – up to a specified date in 2025 – and does not constitute a license for future model training on those works,
- the settlement does not cover potential claims related to AI model outputs, i.e., situations in which AI generates text similar to the original work,
- the settlement applies only to works within the defined class; authors of other works retain the right to bring their own actions.
Who Is in the Class and How the Payout Works?
The class consists of copyright holders (legal and beneficial owners) of works that:
- were downloaded from LibGen or PiLiMi,
- were timely registered in accordance with U.S. rules,
- fall within the period and conditions defined by the settlement.
The payout process is structured around deadlines for:
- opting out of the class for those who do not want to be bound by the settlement,
- filing claims for payment,
- the final fairness hearing, after which payment follows if the court approves the settlement.
In practice, the settlement established an informal standard: the approximate value of compensation per work in such circumstances becomes about 3,000 dollars per book. Although this amount is not binding on other courts, it becomes a strong reference point for future negotiations between authors, publishers, and AI companies.
For the AI industry, the message is clear: systematic use of pirated sources can lead to multi‑billion‑dollar liabilities. For authors and publishers, the message is equally clear: an organized class action can lead to real and meaningful per‑work compensation.
Need legal advice on copyright and AI?
Follow for more legal insights:
Follow for more legal insights:
