Who Owns the Output? The Moot

A scanned book and a stolen one

Anthropic, the company that makes the Claude models, acquired some of the same books twice. First it downloaded more than seven million pirated books from sites such as Library Genesis. Later it bought print copies of millions of books, some of them titles it already had, tore off the bindings, and scanned every page. In June 2025, in Bartz v. Anthropic (opens in a new tab), Judge William Alsup gave the two copies of the same text opposite answers. The scan was fair use. The pirated copy kept in the company’s library was not, and buying the book later did not cure it.

The words were identical, so the difference had to lie somewhere else. Where? Three answers are worth testing: the purpose of each copy, the sale each copy replaced, and what came out of the machine at the end. Each explains part of the case, and each is contested in the OpenAI litigation. The activity puts you in that fight, arguing the training question from the real record.

Is it the purpose of the copy?

Fair use is not a verdict on a copy. It is a verdict on a use. Section 107 asks first about “the purpose and character of the use,” and the Supreme Court reads that factor as a question about what the copier did with the work. In Andy Warhol Foundation v. Goldsmith, the Court said the first factor asks whether, and how far, a use has a purpose different from the original’s. The larger the difference, the more likely the factor favors fair use, and a commercial use that shares the original’s purpose will likely lose it.

Judge Alsup found the purpose of training as different as a purpose gets. A book is written to be read. Anthropic copied books so its models could map the statistical relationships among words and then write new text in reply to new prompts. He called that use transformative, “spectacularly so,” and compared it to a reader who studies books in order to write. That is the core of transformative use: the copy serves a job the original never did.

The pirated library failed the same test. Anthropic’s own plan, the order noted, was to keep everything forever. A permanent library of books the company never paid for serves the purpose a bought book serves, which is to have the book. Some of those copies later fed training. The court still treated building the library as a separate use, and the training that came after did not justify it.

Purpose also explains why another court went the other way. In Thomson Reuters v. Ross (opens in a new tab), a legal research startup copied Westlaw’s headnotes to train a tool that returned judicial opinions in answer to legal questions. Judge Stephanos Bibas found no new purpose. The tool was built to compete with Westlaw, and it generated no new writing of its own. The first factor went to Thomson Reuters. In September 2026 the Third Circuit affirmed (opens in a new tab): both companies used the headnotes to help researchers find opinions, which made Ross’s use “minimally transformative, at best.” The court added that the generative models in Bartz and the OpenAI case were a different matter, because Ross’s tool could not write anything new.

So purpose explains a great deal. It does not explain everything, because a use with a new purpose can still lose on the fourth factor.

Is it the sale the copy replaced?

The fourth factor asks how the use affects the potential market for the work, or its value. Its oldest worry is market harm through substitution: a copy someone takes instead of paying for the original. The pirated library in Bartz was substitution in its plainest form. Every downloaded book was a book Anthropic did not buy, and Judge Alsup found that every factor pointed against fair use for those copies.

Training was harder. The authors argued that models trained on their books would help flood the market with competing writing. Judge Alsup compared that complaint to one against teaching schoolchildren to write well, which could also produce an explosion of competing works. The Act, he wrote, advances original works; it does not protect authors from competition.

Two days later, in Kadrey v. Meta (opens in a new tab), Judge Vince Chhabria rejected that comparison. Teaching a child to write, he said, is nothing like building a product that lets one person generate countless competing works in a fraction of the time. He called this market dilution and said that harm to the market matters more than the purpose of the copying. He predicted that in most cases training without permission would be illegal. Meta still won, because the thirteen authors who sued had barely argued dilution and offered no evidence of it. His ruling, he wrote, stood only for the proposition that these plaintiffs made the wrong arguments.

The judges also split on licensing. Some publishers now sell AI companies licenses to their work, so is an unpaid license a lost sale? Both California judges said no. A copyright owner has no right to a fee for a transformative use, Judge Chhabria explained. Counting the lost fee would tilt the fourth factor toward the owner in every case. Judge Bibas said yes: a potential market for AI training data was enough to weigh against Ross. The Third Circuit agreed. Thomson Reuters had not licensed its headnotes to others, but the evidence showed a market for them as training data taking shape, and Ross’s unlicensed copying took that market from Thomson Reuters.

The fourth factor is where the question is least settled. The Justice Department’s statement in the OpenAI case argues that the dilution theory folds the model’s outputs into its training and misreads the factor. The publishers’ brief relies on Kadrey for the opposite view. When you argue the fourth factor, you are choosing between judges who have already disagreed.

Is it what comes out of the machine?

The third answer looks past training altogether. In Bartz, the authors never claimed that anything Claude wrote copied their books, and the order said the record showed the opposite. That shaped the result: had the models handed the public copies or knockoffs of the books, the judge wrote, it would have been a different case. Output infringement is a different claim, about what a model gives back to a user. It turns on whether an output copies an author’s expression rather than the facts or style it shares with everything else.

The news cases put output at the center of the fight. The publishers point to the words of OpenAI’s and Microsoft’s own executives, and to answers that reproduce or summarize their articles. The products, they argue, substitute for the news. OpenAI answers that nobody owns facts. It says the publishers’ own experts measured word-for-word regurgitation at a tiny fraction of a percent, and that its models are trained to refuse such requests. Both sides agree that a particular output would need its own infringement analysis. They disagree about whether outputs belong in the training argument at all.

The title of this activity asks a last question: who owns what the machine writes? Here the law is firmest. In Thaler v. Perlmutter (opens in a new tab), the D.C. Circuit held that the Copyright Act makes a human being a necessary condition of authorship. A picture its maker attributed entirely to software could not be registered. That is the rule of human authorship. A person who uses an AI tool can still own what the person adds: the selection, the arrangement, the edits. The Copyright Office asks applicants to identify and exclude the rest. In other words, no one holds a copyright in the machine-made part of an output.

What a court can decide from the record

The OpenAI case is at summary judgment. At that stage a court decides a question without a trial, if the record leaves no genuine dispute about the facts that matter. Fair use often gets decided there. In Ross, once the copying itself was undisputed, Judge Bibas treated fair use as a question of applying law to facts, and so a question for the judge. That is why the briefs you read are dense with record citations. A fact that is not in a filing does not exist at this stage, however true it may be.

So the puzzle of the scanned book and the stolen one does not have a single answer. Purpose separated the library from the training, the market sealed the fate of the pirated copies, and the lack of infringing outputs kept the training argument clean. The news publishers are trying to break that last condition, and OpenAI is trying to keep it.

In the moot, you argue one side of whether OpenAI’s copying to train its models is fair use, factor by factor, from five excerpts of the real briefs. Read them with a pen out, marking each fact and its page, then run the Fair Use Tutor and decide which factor you think the case turns on. The judge will ask.