The Developer Chose the Training Data

The claim: using the tool makes the author a pirate

The accusation takes a documented act by a model developer—acquiring books from pirate libraries—and assigns that act to every writer who later uses the company’s product. The first fact does not establish the second. Buying or using a commercial tool does not rewrite history and make the customer the person who assembled its training data.

Reality check: what the court records establish

Major model developers did acquire book collections from pirate sources. In Bartz v. Anthropic, the federal district court record described millions of books downloaded from Library Genesis and Pirate Library Mirror, sources Anthropic knew contained pirated copies. The court separated that acquisition and retention from the later use of books for model training: it found the training use transformative on the record before it, while refusing to excuse the creation of a permanent library from pirated copies. Anthropic later entered a class settlement concerning those pirated books.

Meta’s litigation record likewise described books obtained from LibGen for Llama training. The ruling in Kadrey v. Meta was expressly limited: the court said the plaintiffs in that case had not developed the right record to prevail, not that every use of copyrighted material for training was automatically lawful.

How responsibility is divided

The developer selected or approved training sources, copied the files, built the dataset, trained the model, and controlled what it disclosed about that process. A novelist who later opened a commercial writing tool did not make those upstream acquisition decisions. The author may not know which individual works were present, may have no access to the dataset, and cannot retroactively choose how the model was trained.

That does not end every ethical question. A downstream user can decide whether to financially support a company, what tools to use, what promises to make readers, whether to disclose generated material when required, and whether an output reproduces protected expression. An author who knowingly passes off copied text, violates a platform rule, or lies to customers owns those choices. But “you used a model whose developer used pirated books” is not proof that “you pirated those books,” selected the dataset, or intended the upstream copying.

Writers have always learned by reading

A person can read thousands of books and carry their rhythms, vocabulary, structures, genre conventions, and stylistic lessons into later work. We do not ordinarily treat influence as evidence that the writer copied a particular book. Copyright protects an author’s specific expression; it does not give one writer ownership of every idea, method, word, or general technique another writer learned by reading.

Human learning and model training are not technically identical. A person remembering what they read is not the same process as a company assembling millions of copied files into a commercial training corpus. That distinction matters when judging the developer’s acquisition conduct. But it does not justify abandoning evidence when judging a downstream writer. An author who uses Claude for brainstorming, critique, research support, or revision should be assessed on what they actually publish: whether they copied protected expression, deceived readers, violated a platform rule, or misrepresented their process.

Anthropic is paying for the covered conduct

On July 20, 2026, the court gave final approval to a $1.5 billion settlement covering claims about books Anthropic downloaded from pirate libraries. That is concrete accountability and compensation for the covered works. It does not retroactively make the downloads proper, provide a universal license for future conduct, or answer every unresolved question about AI training.

It does mean the company that made the acquisition decisions is paying to resolve those claims. A present-day author using Claude does not inherit the act of downloading that library or the developer’s resulting liability. The author does inherit responsibility for their own choices: how the tool is used, how much human authorship they contribute, what they disclose when rules require it, and what they put into the marketplace.

Complicity is an argument—not proof of personal piracy

People may argue that using a product creates indirect complicity in its maker’s conduct. That is a moral and consumer-choice argument; state it as one. Do not quietly replace it with a factual accusation that the current author personally stole, uploaded, or downloaded books when there is no evidence they did.

If the complaint concerns the training library, examine the entity that assembled the training library. If it concerns the book now being sold, examine the author’s disclosure, process, output, and representations. Accurate accountability begins by assigning each act to the actor who performed it.