Who owns the words that trained AI? Courts are still figuring it out
A $1.5 billion ruling against Anthropic and a string of unresolved lawsuits show just how far copyright law is from catching up with the technology reshaping publishing.
Key points
- Judge William Alsup ordered Anthropic to pay a $1.5 billion copyright settlement to authors in 2024, but ruled that AI training itself was lawful.
- U.S. Copyright law hasn't been updated since 1976, leaving courts to apply 50-year-old rules to AI.
- A separate ruling found that fully AI-generated works can't be copyrighted under current law.
- Most major AI copyright cases remain in active litigation, with no final resolution in sight.
Every time you use a chatbot like ChatGPT or Claude, you're talking to a system that learned from an enormous library of text: hundreds of millions of books, articles and academic papers scraped from the internet. Most of the authors whose words filled that library never agreed to it. Many don't even know it happened.
The honest legal answer is that nobody knows yet whether it's illegal. That uncertainty has produced a string of partial rulings, each influential, none definitive.
What have the courts actually decided?
The rulings so far point in different directions, and none is final. In one of the first major decisions, Judge William Alsup ordered Anthropic, the company behind the Claude family of AI assistants, to pay a $1.5 billion settlement to a group of writers. That sounds like a win for authors. It wasn't quite that simple.
Alsup ruled that training an AI on copyrighted books is lawful. What he penalised Anthropic for was sourcing those books from illegal pirate websites. The fine was for theft of files, not for the act of learning from them.
In his written opinion, Alsup compared the way a large language model (the technology that reads vast amounts of text and learns statistical patterns from it) ingests books to the way a human writer studies literature before finding their own voice. Attorney Cathy Gellis, who specialises in intellectual property and technology law, told TechCrunch that framing was good news for AI companies: copyright law, she explained, hinges on copying a work, not on reading or consuming it.
A separate case drew a sharper line. When the media company Thomson Reuters sued the legal research firm Ross Intelligence for copying its content to build a competing AI-powered legal search tool, Judge Stephanos Bibas sided with Reuters. His reasoning: if you train an AI on someone else's material specifically to compete with them in the same market, that isn't fair use. We first covered that dispute on 23 August 2026 in our Thomson Reuters v. Ross Intelligence report.
"Fair use" is the legal concept that allows limited, unlicensed use of copyrighted material for purposes like criticism or education. Courts weigh several factors, including whether the new work competes directly with the original.
Does it matter who the AI competes with?
Right now, that appears to be the key question. Attorney Jason Henderson, founder of the IP and Media Practice at JWL International, put it plainly to TechCrunch: if an AI company trains on your work to build a product that directly replaces you in the market, courts are leaning against that. If the trained model does something distinct, courts are more likely to find a way to permit it.
For authors, that leaves a difficult gap. Chatbots can generate entire novels on demand. Whether that counts as market competition in the legal sense hasn't yet been tested to a verdict.
| Case | Ruling | Year |
|---|---|---|
| Anthropic authors' settlement | Training lawful; piracy source penalised ($1.5 billion) | 2024 |
| Thomson Reuters v. Ross Intelligence | Training to compete directly ruled not fair use | 2024 |
| Thaler v. Perlmutter | Fully AI-generated works cannot be copyrighted | 2023 |
What does this mean for writers and readers?
For working authors, the current situation offers little certainty. Their books almost certainly trained at least one major AI model. Compensation is unlikely unless a court rules otherwise, and no such ruling has come. It's worth watching how AI companies source training data: as we reported on 12 August 2026, some firms have been buying and physically destroying books to scan them cheaply, a practice booksellers are already raising alarms about.
U.S. Copyright law was last updated in 1976, long before the internet existed. Until Congress rewrites it or a higher court settles the question, every ruling is provisional.
Common questions
Can an author sue if their book trained an AI?
Yes, and many have filed lawsuits, but winning is hard. Courts have so far focused on how the books were obtained and whether the AI directly competes with the author, not simply on whether the book was used.
If an AI writes a book, who owns the copyright?
Currently, nobody. The 2023 ruling in Thaler v. Perlmutter established that a fully AI-generated work can't be copyrighted under U.S. Law, because copyright requires human authorship.
Will the law change soon?
Congress has held hearings but passed nothing yet. Real clarity will most likely come from a combination of higher-court rulings and eventual legislation, a process legal experts expect to take several more years.
