The 1.5 Billion Dollar Anthropic Settlement and What the Fair Use Line Means for Your Firm
Bindingness: Binding Law · Scope: Federal · Legal
A federal judge drew the clearest line yet between training on books you bought and hoarding books you pirated. The settlement money is being counted now. Here is what that line means for the documents you feed AI.
Part of AI Regulation Tracker, our running brief on the laws reshaping AI at work.
Key takeaways
- The line is acquisition, not output. On the facts of this case, the court praised the training as transformative and still let the piracy claim proceed. How you obtained the material is its own legal question.
- Bought and digitized was fine; pirated and stockpiled was not. Buying print books and scanning them for an internal library counted as fair use. Pulling seven million copies from pirate sites did not.
- The money is real and moving. At least 1.5 billion dollars, with a headline figure of roughly 3,000 dollars per work across about 465,000 to 500,000 books before fees and allocation, with distributions being calculated now.
- The practical question for professionals is provenance. Before you feed a document to AI, you should be able to say where it came from and whether you have the right to use it that way.
What the Court Actually Held
On June 23, 2025, Judge William Alsup issued the first substantive ruling in the United States on how copyright fair use applies to generative AI training. The case, Bartz v. Anthropic, was brought by authors who said the company used their books to build Claude without permission. The decision split into two clean parts, and the split is the whole lesson.
On the training itself, Anthropic won. The court called the use of copyrighted books to train a large language model transformative, and in its own words, spectacularly so. The judge also accepted that buying print books and scanning them into a searchable internal library was fair use, because Anthropic had simply replaced copies it lawfully owned with more convenient digital ones, without creating new copies or redistributing anything.
On the piracy, Anthropic lost. The court found that the company downloaded more than seven million books from known pirate libraries such as Library Genesis and the Pirate Library Mirror, then kept them to build a permanent central library of, in the company's framing, all the books in the world. Judge Alsup refused to bless that. His line was direct: you cannot bless yourself by saying you have a research purpose and then go take any book you want. The piracy claim was allowed to head toward trial.
Buy the book and you may be free to learn from it. Steal the book and the learning does not launder the theft.
Facing that exposure, Anthropic settled. The company agreed to pay at least 1.5 billion dollars plus interest, described by plaintiffs counsel as likely the largest copyright recovery ever. The headline math works out to roughly 3,000 dollars for each of the eligible books, a class that lands somewhere between about 465,000 and 500,000 works after duplicates and ineligible titles are removed, though the final per work figure depends on fees, costs, and the allocation formula approved by the court.
Where the Money Stands in June 2026
The settlement has moved through the courts in stages rather than one signature. Anthropic funded an early installment after preliminary approval, and the agreement schedules the rest in further payments tied to approval anniversaries. The claims rate was high. By early 2026 the settlement administrator reported that more than 91 percent of eligible works had been claimed.
The settlement administrator was set to calculate individual distributions by June 11, 2026, within 80 days of the claims deadline. That date is when the math gets done, not when checks clear. At a hearing on May 14, 2026, the judge overseeing approval declined to grant final approval on the spot and asked for more detail on fees and lead plaintiff payments. Actual payouts to class members follow final approval and the resolution of any appeals. So as of late June 2026, the figures are being finalized and the precedent is set, even as the last administrative steps play out.
This was a fight about training data, but the principle is about provenance
You are not building a frontier model. You are pasting a contract, a client file, or a market report into a chat window. The court did not rule on that. But the reasoning, that lawful acquisition is its own requirement separate from clever use, is exactly the question a careful professional should ask before every upload.
Fair Use Versus Infringement: The Line the Court Drew
The ruling is easiest to hold onto as a contrast. The same activity, copying a book into a computer, fell on opposite sides of the line depending entirely on how the book was obtained and what was done with it.
| What Anthropic did | How the court treated it |
|---|---|
| Trained the model on copyrighted books | Fair use. The court called it transformative, spectacularly so. |
| Bought print books and digitized them for an internal library | Fair use. A format shift of copies it lawfully owned. |
| Downloaded millions of pirated books and stockpiled them | Not fair use. The piracy claim proceeded toward trial and drove the settlement. |
Read across those rows and the message is plain. The model learning from a work was forgiven. The way the work was acquired was not. A research goal at the end of the pipeline did not cleanse a pirated source at the start of it.
What Is Safe to Feed AI: A Professional's Decision Checklist
Anthropic's 1.5 billion dollar problem came from building a model on a pirated library. Your risk is different in scale but not in principle. The same reasoning that governs a library of seven million books applies to the single market report or contract you are about to upload, because the court's line was about how the material was obtained, not how big the pile was.
You can borrow the court's logic without being a copyright lawyer. Before you paste a document into an AI tool, walk it down this checklist. It separates the acquisition question, where this case lives, from the confidentiality question, which your professional duties govern. If a document fails a line, stop and route it to the safer column or to a human review first.
Run every document through these five questions before it touches an AI tool
- 1. Provenance. Do I know where this document came from and that it was lawfully obtained? If it arrived from a scraped dump, a pirate source, or an unknown forward, treat it as off limits until you confirm.
- 2. Rights. Do I have the right to use this material this way? Your own work product, your firm's templates, public records, and content you are licensed to use are the safe core. A third party's copyrighted material you merely possess is not the same as material you may copy into a tool.
- 3. Confidentiality. Is this client or privileged information, and is the tool approved to receive it? Acquisition being clean does not make the upload clean. A lawfully obtained client file can still breach confidentiality if it goes into an unapproved tool.
- 4. Training and retention. Does this tool train on my inputs or retain them, and have I checked the current terms? Prefer tools that contractually do not train on your content for anything sensitive, and confirm the setting rather than assume it.
- 5. Verification and ownership. Will a qualified human review the output, and do I own the result before it reaches a client or a court? AI drafts; you remain responsible for accuracy, citations, and the final work.
Most day to day professional work passes this check comfortably. Your own memos, your firm's prior briefs, a contract your client sent you, public filings, and material you are licensed to use are the safe foundation. The documents that should give you pause are the ones with murky origins or third party copyright you never cleared, and the client material that belongs only in an approved, confidential tool.
Attorneys who want the deeper treatment of confidentiality, verification, and defensible AI use can start with The Leveraged Attorney.
Frequently Asked Questions
Did the court say AI training on copyrighted books is legal?
In this case, yes, with a sharp caveat. Judge Alsup found that training an AI model on copyrighted books, and digitizing books the company had lawfully purchased, was fair use because the use was transformative. He did not extend that to the separately pirated copies. This is one district court ruling, not a Supreme Court rule, and other AI copyright cases are still moving, so the law is not finished.
What exactly made Anthropic pay 1.5 billion dollars?
Not the training. The exposure came from how the books were acquired and stored. The court found Anthropic downloaded more than seven million books from pirate sites and kept them in a permanent library, which it would not treat as fair use. Facing trial on that piracy claim, the company settled for at least 1.5 billion dollars, roughly 3,000 dollars per eligible book.
Does this case mean I cannot paste documents into Claude or ChatGPT?
No. The case is about a company building a training library from pirated books, not about a professional using an AI tool on documents they lawfully hold. The transferable lesson is provenance: be able to say where a document came from and that you have the right to use it. Your own work, licensed material, public records, and client files you were given are the safe core, subject to your confidentiality duties.
When will authors actually be paid?
The settlement administrator was scheduled to calculate distributions by June 11, 2026, but that is the calculation step, not the payout. At a May 14, 2026 hearing the court asked for more detail before granting final approval. Payments to class members follow final approval and any appeals, so as of late June 2026 the amounts are being finalized while the last court steps proceed.
What is the single most useful rule to take from this?
Separate two questions before every upload. First, acquisition: did I obtain this lawfully and do I have the right to use it this way? Second, confidentiality: is this client or privileged material, and is the tool approved for it? This case is about the first question. Your professional duties govern the second. Clean on both is your green light.
Related: The confidentiality guide for attorneys using AI, why AI citations get lawyers sanctioned, and where AI is safe for legal research versus drafting.
Sources: Bartz v. Anthropic, N.D. Cal., summary judgment order of June 23, 2025; The Authors Guild, what authors need to know about the Anthropic settlement and the 91.3 percent claims update; CBS News and PBS NewsHour reporting on the 1.5 billion dollar settlement and the roughly 3,000 dollar per book figure; Jones Walker LLP AI Law Blog analysis of the fair use ruling and the data acquisition lesson; Courthouse News Service and Publishers Weekly on the May 14, 2026 approval hearing. Court rulings and settlement terms change and may be appealed; confirm against current filings and your own counsel before relying on this page. This page explains the case; it is not legal advice.