A federal court in San Francisco just handed down a ruling that is shaking up authors and AI users alike. Judge William Alsup ruled that Anthropic, the AI company behind the Claude chatbot, did not break the law by training its models on copyrighted books—at least not when it comes to how the books were used. But how Anthropic got those books is another story, and may subject Anthropic to liability when a jury takes up that issue in December.
What the Court Said
In his opinion issued yesterday, Judge Alsup ruled training AI models on books qualifies as “fair use” under U.S. copyright law. According to the court, the AI was not copying the books word-for-word. Instead, it was learning from them the way a writer might read many books before writing something new. That is, the judge determined, what the law calls “exceedingly transformative,” making it fair use under Section 107 of the Copyright Act. The court also ruled that Anthropic’s digitization of the books it had purchased in print was also fair use.
That’s a big win for tech companies. It’s the first court decision in the country to clearly say that using books to train AI is legal under fair use—so long as the books are used to create something new and not just copied.
Where Anthropic Got in Trouble
But that’s not the end of the story. The judge went on to rule that fair use does not extend to making copies of non-purchased copyrighted work. The court ruled that Anthropic’s method of collecting those non-purchased books—by downloading them from pirated sources like Books3 and Library Genesis—was not legal. Storing over 7 million pirated books in a “central library” was copyright infringement, plain and simple, and not protected under fair use.
Judge Alsup wrote, “Anthropic had no entitlement to use pirated copies for its central library” and pointed out that buying a book after stealing it doesn’t make the earlier theft okay.
What Happens Next
A trial is set for December to figure out how much money Anthropic might owe for copyright damages. If the court finds the company’s infringement was “willful,” Anthropic could face up to $150,000 in damages for each work it copied.
Why This Matters
Assuming this ruling is not appealed and overturned, this ruling provides AI companies with a roadmap on how to implement best practices in AI training. For right now, training AI on copyrighted material you own appears to be fair use—but training AI on pirated material you do not own is still a big no no. That means tech companies cannot just download pirated content and claim fair use later. They will need to license books or get them through other legal sources.
The ruling could also have an immediate impact on other large AI players like OpenAI and Meta, who are facing similar lawsuits. This case sets the tone and framework for how courts might treat these AI training cases moving forward.
Best Practices
If you are building AI tools, now is the time to double-check where your training data comes from. Just because you are creating something new does not mean you can ignore how you got there. As this case shows, courts are watching—and they are ready to act when AI companies cross the line.
Recent Comments