US court rules AI training on copyrighted books lawful but penalises pirated data sources

A recent US court ruling clarified that training AI models on copyrighted books does not automatically violate copyright law. The decision frames AI training as analogous to a human reading and learning from published works, rather than copying them. This interpretation aligns with existing fair use principles, which permit the use of copyrighted material for transformative purposes. However, the ruling does not eliminate legal risks entirely, as courts may still scrutinise whether AI outputs compete with original works in the marketplace.

The ruling also imposed a $1.5 billion penalty on Anthropic, but not for the act of training itself. The fine targeted the company’s use of pirated books sourced from illegal shadow libraries. This distinction is critical for engineers: the legality of training data depends on how it was obtained, not just how it is used. Companies must now ensure their datasets are legally sourced, which may require costly audits and provenance tracking. Failure to do so could expose them to similar penalties, regardless of the training process’s compliance with copyright law.

The legal landscape remains unsettled because copyright law has not been updated to address AI-specific challenges. Courts are interpreting 50-year-old statutes to decide cases involving generative models, leading to inconsistent rulings. For example, a separate case found that training an AI to compete directly with a copyrighted work was not fair use. This creates uncertainty for engineers, as the same training process could be deemed lawful or unlawful depending on the AI’s output and its market impact. Until legislation or higher court rulings provide clearer guidance, companies must weigh the risks of litigation against the benefits of using copyrighted material.

Fair use determinations in AI cases currently hinge on whether the model’s outputs compete with the original works. If an AI generates content that displaces demand for copyrighted books, courts may rule against the training process. However, if the AI’s outputs are sufficiently transformative or serve a different market, the training may be deemed lawful. This standard is difficult to apply in practice, as generative models can produce outputs that both compete with and complement original works. Engineers must consider these nuances when designing models, as the legal risks may vary depending on the use case and deployment context.