Anthropic Wins: Training AI on Books Is Fair Use, but Piracy Isn't

On June 24, 2025, Judge William Alsup separated training use of books from acquisition and retention of pirated copies. The original source supports the documentary core of the event; the order resolved distinct issues differently and did not turn piracy into fair use.

The decision comes in Bartz v. Anthropic, a class-action lawsuit brought by authors against the company behind Claude. It's the first time a court has credited the argument AI giants have been making for years: that training a model on copyrighted works falls under fair use, the legal doctrine that allows protected material to be used without the rights holder's permission under certain conditions.

What the ruling actually means

Fair use is one of the slipperiest concepts in U.S. intellectual property law, and it hasn't been updated since 1976 — long before the internet existed, let alone the notion of training datasets for generative AI. Courts weigh several factors when deciding whether a use is legitimate: what the work is being used for (parody or education tend to get more latitude), whether there's direct commercial gain from the reproduction itself (you can write fan fiction based on a copyrighted saga, but you can't sell it), and above all, how transformative the result is compared to the original.

Alsup found that Anthropic's training of its models on published books was transformative, in line with the standard reasoning applied to language models: the idea that these systems don't simply reproduce the works, but use them to generate new content. That's the exact argument companies like Meta have made in their own defenses against similar lawsuits, and until now it was far from clear whether any judge would back it with a ruling of this weight.

The decision doesn't bind other judges in the dozens of pending lawsuits against OpenAI, Meta, Midjourney, Google, and others. But it does set a precedent: for the first time, a court has put in writing that training itself can be shielded by fair use, giving AI companies a real legal foothold against authors, artists, and publishers who have spent years trying to block the practice in court.

The other half of the case: piracy, no ambiguity

Alsup's ruling doesn't close the Bartz v. Anthropic case. The plaintiff authors didn't just challenge the model's training — they also zeroed in on how Anthropic obtained and stored their works. According to the lawsuit, the company set out to build a "central library" of "all the books in the world" to keep "forever." The problem is that millions of those copyrighted books were downloaded for free from pirate sites, something the law leaves no room to doubt: it's illegal.

Here the judge draws a hard line between the two issues. The fact that training on those books qualifies as fair use doesn't launder the method Anthropic used to acquire them. Alsup has ordered a separate trial on the nature of that "central library" and the damages it may have caused.

In his ruling, the judge put it plainly: "We will have a trial on the pirated copies used to create Anthropic's central library and the resulting damages," he wrote, as reported by TechCrunch. He added a clarification that makes clear buying the book afterward doesn't erase the original offense: "That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for theft but it may affect the extent of statutory damages."

Why this distinction matters for the rest of the industry

The ruling draws a line that will likely shape the rest of copyright litigation against AI companies: how a text is used to train a model is one question, and how that text was obtained in the first place is an entirely different one. A company can have solid fair use arguments for training and, at the same time, face legal and financial liability for turning to piracy to gather the material.

For the authors, artists, and publishers who have filed similar lawsuits against other companies, this is a double blow: they lose the core argument over whether training itself is illegal, but they retain a very specific line of attack if they can prove the material was obtained through piracy. For AI companies, the message is just as clear: where training data comes from can matter as much as — or more than — how it's later used.

The trial over Anthropic's "central library" will now determine the real scope of the damages, in a case that already stands as the first serious precedent in the legal battle between the AI training industry and the publishing world.

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 order resolved distinct issues differently and did not turn piracy into fair use. 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 read each use of a work separately instead of summarizing the whole case as a victory, 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 read each use of a work separately instead of summarizing the whole case as a victory. 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.