AI Policy

A Court Just Priced a Pirated Book at 3,000 Dollars

On 20 July 2026 a federal judge gave final approval to Anthropic's 1.5 billion dollar copyright settlement: 3,000 dollars per work across roughly 500,000 works. Training data provenance now has a published price.

AI PolicyBy Servola Tech Desk2026-07-213 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

A Court Just Priced a Pirated Book at 3,000 Dollars

Key takeaways

A federal judge in San Francisco gave final approval on 20 July 2026 to Anthropic's 1.5 billion dollar settlement with authors, paying 3,000 dollars for each of roughly 500,000 works, which means training data acquired the wrong way now carries a published per-item price rather than an unknown one.
The ruling underneath the settlement did not say training on books is unlawful. Judge William Alsup had held that training on lawfully obtained text was fair use and that downloading the books from pirate sources was not, so the money buys forgiveness for acquisition and not for use.
Ask your model vendor how its training corpus was acquired, not whether it was licensed. Those are different questions, and only the first is the one a court has now put a number on.
The settlement binds a class and not the world. Objections were filed that it excludes some copyright owners, and rightsholders outside the class keep their claims, so a vendor pointing at this settlement as proof it is clear has not answered the question you asked.

The signature that closed the biggest one

Judge Araceli Martinez-Olguin signed on Monday. On 20 July 2026 the United States District Court for the Northern District of California gave final approval to a 1.5 billion dollar settlement between Anthropic and a class of authors and publishers, the largest settlement on record in American copyright law. Payments of 3,000 dollars per work, across roughly 500,000 works, can now begin.

The case did not reach her in the ordinary way. Judge William Alsup had presided, issued the substantive rulings and granted preliminary approval last September, and then retired. Martinez-Olguin inherited a settlement rather than a trial. She also awarded plaintiffs' counsel more than 101 million dollars in fees against the 187.5 million they had requested, and approved over objections from authors who argued the sum was too small, the fees too large, and the class boundary drawn to exclude people who belonged inside it.

What was illegal was not what most people think

The distinction is the whole case. Alsup held that training a large language model on lawfully obtained copyrighted text was fair use. He also held that Anthropic's downloading and retention of millions of books from pirate sources was not. The company settled rather than appeal. So the 1.5 billion dollars is not the price of teaching a model to read. It is the price of how the library was assembled.

That is a narrower holding than the headlines suggest and a more uncomfortable one for buyers. A fair-use defence protects the act of training. It does nothing for the act of acquisition, and acquisition is the part that leaves records: downloads, transfers, invoices, storage. A vendor can be entirely right about fair use and still carry the liability that actually got priced this week.

3,000 dollars is now the number in the room

Before Monday, provenance risk in an AI contract was a quantity nobody could size. Now there is a published figure from an approved American settlement: 3,000 dollars per work, across a class of roughly half a million. Any lawyer negotiating an indemnity, any insurer writing a policy and any counterparty arguing about a liability cap will reach for that number, because it is the only one with a judge's signature on it.

For an operator this changes what to ask rather than what to fear. The useful procurement question is not whether your vendor licensed content. It is how the training corpus was acquired, item by item, and whether the vendor can evidence it. Ask for the answer in the contract, with an indemnity that survives a claim brought by someone who was never in an American class action. A vendor that will not write it down has told you where it thinks the risk sits.

Why this does not settle the European question

A settlement binds the parties to it. Rightsholders outside the certified class keep their claims, which is precisely what several objectors argued about, and no European rightsholder was resolved by a Californian class action. The European route runs through the text and data mining exception, which turns on lawful access and on whether a machine-readable reservation of rights was present when the material was collected. That is an acquisition test as well.

So the same question follows the model across the Atlantic in a different accent. Was the material lawfully accessed at the moment it was collected, and can anyone still prove it. European buyers of American models should assume that a United States settlement retires American class claims and nothing else, and should hold their own provenance warranty accordingly. The price of the answer is now on the record. The answer itself still is not.

Frequently asked questions

Does this mean training AI on books is illegal now?

No, and the ruling underneath the settlement says close to the opposite. Judge William Alsup held that training a model on lawfully obtained copyrighted text was fair use. What he held unlawful was downloading and keeping millions of books from pirate sources. The settlement resolves that second question with money and leaves the first as decided law in that court.

We use an American model. Are we exposed?

Not by this settlement directly, because it binds a class of American authors and publishers against one company. Your exposure runs through your contract. Ask your vendor how its training corpus was acquired and whether it will indemnify you against provenance claims, including claims from rightsholders who were never part of a United States class action. The answer belongs in the agreement rather than in a sales call.

Why did the attorneys receive 101 million dollars?

Plaintiffs' counsel requested 187.5 million dollars and the court awarded more than 101 million of it. Fee awards in class settlements are set by the court rather than agreed between the parties, and reductions of this scale are common where a judge considers the requested share disproportionate to the fund. Several objecting authors had argued the fees were too high and the settlement too small.

The court did not price what a model learned. It priced how the books were obtained, and that is the bill your contract has to name.

Anthropiccopyrighttraining dataclass actionprovenancetext and data mining

Microsoft Trained Its Sellers to Call Claude Slower

Microsoft trained its sellers to call Claude slower and less accurate inside Office. A separate report says it is swapping the model out to cut cost. Read the ranking accordingly.

4 min read

Energy

A Bitcoin Miner Just Signed a $19bn AI Power Lease

TeraWulf, a former bitcoin miner, has signed a 20-year, $19 billion lease to supply Anthropic with about 401 megawatts of AI capacity in Kentucky. The scarce asset in the AI build-out is powered land, and the capacity operators will rent in 2028 is being locked up now.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.