AI
Copyright Law's AI Gray Zone: Why Training on Books Sits in Legal Limbo
Major AI models learned from millions of books without author consent, but recent court rulings suggest the practice may be legal—complicating a decades-old copyright framework.
TechCrunch — AI
Modern large language models like ChatGPT, Gemini, and Claude have absorbed vast troves of books, articles, and web content to learn how to generate text. Most authors never consented to their work being used this way, and many see the technology as a threat to their craft. Yet the legal status of this training remains murky, caught between 1976 copyright law and 2020s-era machine learning.
The Landmark Ruling That Surprised Everyone#
Last year, Judge William Alsup handed down a decision that seemed contradictory: he ordered Anthropic to pay $1.5 billion in damages to writers whose works trained its AI systems—yet he actually ruled that training itself was not unlawful. Instead, he penalized the company for obtaining those books through piracy, from illegal shadow libraries.
Alsup compared how an AI model absorbs trillions of words to how a writer learns by studying literature: both involve reading to improve craft, not copying to reproduce. Copyright law, the judge suggested, protects the copying of works, not the act of consuming them.
For AI companies, this framing is a strategic win. A $1.5 billion fine is manageable for a sector projecting valuations near $200 billion. "I think it is generally good news for AI training," said Cathy Gellis, an intellectual property attorney, noting that Alsup treated model training more like reading than replication.
Fair Use—The Ever-Shifting Standard#
Whether AI training qualifies as "fair use" under copyright law hinges on several factors: whether the use is transformative, how much of the original work is used, and whether it damages the original creator's market. Recent court decisions have been inconsistent, leaving the tech industry uncertain.
A 2024 case offers a contrasting example. Thomson Reuters sued research platform Ross Intelligence for training its competing legal AI on Reuters' proprietary content. Judge Stephanos Bibas ruled this was not fair use, since Ross's purpose was direct competition: the AI system "does not have a 'further purpose or different character'" than the original publication.
Yet authors arguing that chatbots compete with them by generating synthetic books have not yet succeeded with this logic in court. Senior attorney Jason Henderson notes the pattern: courts tend to reject fair use claims when training aims to build a direct competitor, but accept them when the application is non-competitive.
