Is it legal to train AI models on copyrighted books?
A deep dive into US copyright law after a judge rules Anthropic's AI training is lawful while slapping a $1.5B penalty for pirated shadow libraries.
Stock photo for illustration only, not from the actual event
- Judge Alsup ruled Anthropic's LLM training lawful as it mimics human reading.
- Anthropic was ordered to pay a $1.5 billion settlement over pirated online books.
- US copyright law remains unchanged since 1976, forcing courts to adapt old rules.
AI models powering chatbots like ChatGPT, Gemini, and Claude are trained on massive datasets featuring hundreds of millions of books, articles, and academic papers without the knowledge or consent of authors. While this practice appears unlawful at first glance, the legal reality is far more intricate.
Cathy Gellis, an intellectual property and technology attorney, explained to TechCrunch that this legal landscape involves complex mechanisms and deeply held feelings from both sides of the debate.
Last year, Judge William Alsup ordered Anthropic to pay a massive $1.5 billion settlement to authors whose works trained its AI models. On the surface, this seemed like a win for writers, but the judge actually ruled that training AI on those works was entirely lawful. The penalty was issued specifically because Anthropic pirated those books from illegal online shadow libraries.
Stock photo for illustration only, not from the actual event
"Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different"
Judge William Alsup
Judge Alsup compared an LLM ingesting trillions of words to a writer studying literature. Gellis considers the ruling favorable for AI companies, especially given that Anthropic projects roughly $200 billion in annual revenue by 2028.
From an analytical perspective, current copyright frameworks were designed long before modern digital intelligence existed. Courts relying on traditional fair-use doctrines must now evaluate whether training machine learning models constitutes transformative use or market substitution. Because generative models learn underlying patterns rather than direct replication, establishing clear legal boundaries remains one of the tech industry's most pressing challenges.
Jason Henderson, founder of the IP & Media Practice at JWL International, noted that legal experts are deeply concerned because judicial interpretations vary widely. Courts frequently examine whether the AI training serves a direct competing purpose, citing cases like Thomson Reuters versus Ross Intelligence where training on proprietary data to build a competing legal platform was ruled not to be fair use.
Furthermore, cases like Thaler v. Perlmutter establish that 100% AI-generated works cannot be copyrighted, introducing difficult questions about attribution and AI assistance. With most AI companies tied up in ongoing litigation, a definitive legal solution remains far off.
Source: TechCrunch
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment