Anthropic case highlights shifting AI copyright rules

Last year, Judge William Alsup ordered Anthropic to pay a massive $1.5 billion copyright settlement to a group of writers. While this seemed like a blow to the AI developer, the judge actually ruled that training large language models on copyrighted works is lawful. The massive penalty was not for the training itself, but because Anthropic obtained the books from illegal online shadow libraries. Alsup likened the training of models like Claude to a human writer studying literature to create something new.

For AI practitioners, this distinction is crucial. A $1.5 billion fine is relatively minor for a company like Anthropic, which projects roughly $200 billion in annual revenue by 2028. Legal experts note that copyright law, which has not been updated since 1976, relies heavily on the concept of fair use. Under these decades-old guidelines, courts look at whether the AI training is transformative and whether the final product directly competes with the original source material.

This competitive aspect was central to another major case, where Thomson Reuters sued Ross Intelligence. In that instance, Judge Stephanos Bibas ruled against Ross, stating its use was not transformative because it built an AI-based legal platform designed to compete directly with Reuters. For developers building tools like ChatGPT or Gemini, this suggests that training on proprietary data to build a direct competitor carries high legal risks, whereas training to create entirely different creative works may find safer legal harbor.

Beyond training inputs, the industry also faces questions regarding AI outputs. In Thaler v. Perlmutter, a court ruled that entirely AI-generated work cannot be copyrighted. This leaves developers and creators in a gray area regarding how much human assistance is required to secure intellectual property rights. As litigation continues, AI companies must carefully monitor these early rulings, as they will dictate how training datasets are sourced and how synthetic content is commercialized.