Whether an AI company may train a model on copyrighted books remains an unsettled legal question, and recent US rulings have produced a distinction that is easy to miss: training can be treated as lawful even when the way a company obtained the books is not.
TechCrunch’s review of the issue shows why neither authors nor model developers have received a simple answer. Courts are applying a copyright statute last updated in 1976 to systems trained on enormous collections of writing, while different cases turn on how material was acquired, what the resulting product does and whether it competes with the original market.
Training and acquisition can produce different judgments
One prominent case involved Anthropic and a group of writers. Judge William Alsup ruled that Anthropic’s use of books for model training was lawful, comparing the process with a writer studying literature to create something different. The company nevertheless faced a $1.5 billion settlement because the books had been obtained from unauthorized online shadow libraries.
That outcome separates two actions that are often discussed as if they were the same. A court may view the transformation performed during training as permissible while finding that downloading pirated copies infringed copyright. Cathy Gellis, an attorney specializing in intellectual property and technology, told TechCrunch that the training portion of the decision was generally favorable to AI companies.
Her interpretation rests on copyright law’s focus on copying rather than merely using, reading or experiencing a work. Authors may still see their books contributing to tools that threaten their livelihoods, but moral concern and legal infringement do not always map neatly onto each other.
The Anthropic decision is influential, not universal. Gellis cautioned that early rulings can be reversed or contradicted as other courts address different facts. Much of the litigation involving AI developers remains unresolved, so companies and rights holders are making decisions before a stable body of precedent exists.
Fair use depends on purpose and market impact
Many disputes turn on fair use, which can permit copyrighted material to be used without direct authorization. Courts consider factors including the purpose and character of the new use, how much of the original was taken and the effect on its market. A use described as transformative has a stronger argument than one that simply replaces the source.
Attorney Jason Henderson told TechCrunch that outcomes have tended to become less favorable when training uses another company’s property to build a directly competing product. He pointed to Thomson Reuters’ case against Ross Intelligence, which copied legal content to create an AI research platform serving a similar function.
Judge Stephanos Bibas concluded that Ross’s use was not transformative because it lacked a sufficiently different purpose or character. Authors could argue that chatbots compete with them when they generate synthetic books based on training material, but TechCrunch reports that this argument has not yet prevailed in court.
The difference between a general-purpose model learning patterns across many works and a product reproducing the commercial function of a specific database may therefore matter. It does not create an automatic rule: fair-use decisions remain fact-specific, and courts have not reached a final consensus for AI training.
Generated works raise a separate ownership problem
The legality of training data is also distinct from whether AI output can receive copyright protection. In Thaler v. Perlmutter, a court ruled that a work generated entirely by AI could not be copyrighted. That decision raises practical questions about mixed works created through both human effort and automated assistance.
Gellis compared the issue with an author using spell-check software, where the tool does not become the owner of the novel. More advanced systems make the boundary harder to identify: courts and creators may need to determine how much human direction, selection or revision is required before a work qualifies for protection.
These questions are developing on separate legal tracks. A company might have a defensible fair-use argument for training yet still be liable for acquiring pirated copies. An AI-assisted work might receive protection for its human contribution while a fully automated output does not.
For now, the most accurate answer is conditional. The source, purpose, competitive effect and degree of human authorship can change the analysis. Current rulings are already shaping company behavior, but they have not resolved whether every major form of book-based AI training is lawful.
