Anthropic proposes $1.5bn settlement over pirated books

On September 5, 2025, the parties filed a proposed settlement of claims over books obtained from pirate libraries. The original source supports the documentary core of the event; the proposal still required approval and did not itself grant a payment or final right.

That works out to approximately $3,000 per book. The plaintiffs’ lawyers have described it as the largest known copyright recovery to date, but its legal scope is narrower than the size of the payout suggests: it does not establish that training an artificial intelligence system on copyrighted works is illegal. Document supporting the figure.

The issue was how Anthropic obtained the books

The lawsuit was filed by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson. They accused Anthropic of downloading millions of books from pirate repositories such as LibGen and PiLiMi to build an internal library and train its language models.

In June, U.S. District Judge William Alsup separated two issues that are often conflated in the debate over generative AI.

On one hand, he found that using books to train models could constitute fair use, the U.S. copyright exception, because the process was sufficiently transformative. He also upheld Anthropic’s digitization of legally purchased copies. The company had bought printed books, removed their bindings and scanned them.

On the other hand, the judge rejected the idea that the purpose of training an AI system justified downloading pirated copies and keeping them in a corporate library. The issue still to be decided at trial was not simply what Claude did with the texts, but how Anthropic obtained the copies.

That distinction explains the settlement. It does not create a general obligation to pay a license fee for every work used to train AI. It does, however, put a very high price on obtaining data from illicit sources when legal avenues existed to buy or license it.

A trial could have threatened the company’s viability

The case was scheduled to go to trial in December. U.S. law allows damages of up to $150,000 per work when infringement is found to be willful. Applied to millions of files, that scale opened the door to potential liability in the hundreds of billions of dollars and, in the most extreme scenario, close to $1 trillion. Document supporting the figure.

Anthropic acknowledged in court the enormous pressure that exposure placed on it to reach a settlement. The company has avoided a risk that would have been difficult to absorb even for one of the best-funded AI companies.

The payment is significant but manageable relative to Anthropic’s size. The company announced this week that it had raised $13 billion in a funding round that valued it at $183 billion. It also says it is on track to generate at least $5 billion in revenue over the next 12 months. The settlement represents more than a tenth of the newly raised capital. Document supporting the figure.

In addition to the financial payment, Anthropic will have to destroy the files downloaded from pirate libraries and any copies covered by the settlement that remain under its control. The agreement therefore focuses on where the books came from, not on banning Claude or undoing the training of its models.

An economic benchmark, not a legal precedent

A private settlement does not carry the same weight as a judgment after a trial. Other courts are not required to apply the $3,000-per-work figure or resolve different cases in the same way. Document supporting the figure.

Even so, the amount could become a reference point for future negotiations. Meta is facing a lawsuit over its use of books from LibGen, while OpenAI and Microsoft are involved in several intellectual-property disputes. Other cases are also still pending that center on a different issue: models capable of generating text or images that are too similar to protected works and characters.

The settlement could speed the creation of markets for training data. Major publishers and media companies have already signed licensing deals with AI firms, but those agreements are typically confidential and do not provide a comparable public price. The $3,000 per book in this case is not a market rate: it includes the cost of resolving an allegation of piracy and eliminating the risk of far larger legal damages. Document supporting the figure.

That distinction matters especially for smaller companies. Buying, scanning or licensing entire collections requires resources that large labs can more easily afford. If the industry adopts stricter controls on data provenance, legal compliance could become another barrier to entry.

The judge’s approval is still pending

The settlement now needs preliminary approval from the court and, later, final ratification. The parties must also finalize the list of covered works and arrange the distribution of payments among rights holders.

If approved, the Bartz v. Anthropic case will leave the industry with a practical rule: the fact that training may qualify as fair use does not make every way of collecting books legitimate. For AI developers, documenting where each dataset comes from is becoming just as important as showing what it is used for.

Turning the headline into a check

The document heading says what exists: complaint, motion, order, proposal or judgment. Each word changes what a reader may do. An allegation is unproven; a proposed settlement is unapproved; an order may resolve only one motion. the proposal still required approval and did not itself grant a payment or final right. Before extracting a rule, locate the operative text, date and affected parties.

Break the case down by conduct and claim. Obtaining a copy, retaining it, training with it and generating an output are different acts; a court may treat them differently. The evidence and procedural stage matter too. Summarizing everything as 'won' or 'lost' erases the very boundary that makes the ruling usable.

To assess how to distinguish a proposal, preliminary approval and judgment before acting, build a timeline: filing, response, ruling, possible appeal and pending conditions. Then ask which remedy was requested and which was granted. Maximum damages, a proposed fund and money received are not equivalent either. Dates prevent an early document from being credited with what happened later.

What the record must preserve

A United States ruling does not automatically become a universal rule. Jurisdiction, governing law, facts and parties limit its reach. An article can teach document reading but cannot replace advice on a particular case. The useful skill is knowing which question to take to a professional and which page supports each premise.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to distinguish a proposal, preliminary approval and judgment before acting. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.