Anthropic AI Copyright Ruling is a BIG Deal: Fair Use Wins, Piracy Loses

A copyright ruling in the case Barts versus Anthropic draws a sharp line for AI training: model development can qualify as fair use when it is genuinely transformative, but downloading books from pirate sites does not earn the same protection. The decision, issued by Judge William Alup, matters because it offers a workable framework for future AI copyright disputes—one that turns on how training data is obtained, not just what the model does with it.

Alup’s reasoning treats AI training as “quintessentially transformative.” The judge likened the process to how humans read texts and then create new ones—absorbing information and generating outputs that are meaningfully different from the originals. That framing is important because it directly addresses a common legal objection: that training is just copying. By grounding fair-use analysis in transformation, the ruling gives AI companies a conceptual foundation for arguing that training changes the underlying work in a way that goes beyond mere reproduction.

Yet the ruling simultaneously condemns the data acquisition method. The court distinguished between training on books and the earlier step of obtaining those books through piracy. Anthropic had previously trained models using content sourced from Library Genesis and other dubious sites, including for earlier versions of Claude. That choice, the judge made clear, cannot be treated as a free pass.

The case’s practical turning point came in 2024, when Anthropic shifted toward legal procurement. The company hired Tom Turvy, formerly head of Google’s book scanning project, and tasked him with obtaining books through lawful channels. Anthropic then spent millions—described as a significant share of training costs for its new Sonnet and Opus models—buying physical books, often secondhand, and digitizing them by removing them from bindings and scanning them. Although the physical books were destroyed in the process, the court ruled that the resulting digital copies were legitimate fair use because the books were acquired lawfully.

This “Solomon’s choice” approach—fair use for transformative training, but liability for theft—signals how judges may evaluate future AI cases. The decision also notes Anthropic’s financial capability: if the company could afford to buy books later, it could have purchased them earlier. That observation may not eliminate liability, but it could affect the extent of statutory damages.

For authors, the ruling offers a limited but meaningful reassurance. Even if AI outputs are argued to be fair use, the court’s expectation that companies pay for access supports a more sustainable equilibrium for the creative economy. The message is not that authors will automatically win every argument about AI, but that scraping and stealing work without compensation is less likely to be tolerated.

The decision also lands amid other active lawsuits, including Kadre versus Meta over training Llama on Books 3, and potential disputes involving image generation models that may try to extend transformative-use reasoning from text. Broader uncertainty remains because courts differ across jurisdictions, with circuit splits affecting platform liability standards. Still, the Alup framework—transformative use plus scrutiny of lawful data acquisition—provides clearer guidance for where AI copyright fights may head next.