A U.S. federal judge ruled that using lawfully obtained books to train Anthropic’s Claude AI models counts as fair use.
But don’t pop the champagne just yet. The same court made it very clear: pirated books are a whole different story. And Anthropic allegedly used over 7 million of them.
What Anthropic Is Saying
Anthropic hasn’t released a flashy victory statement, but the ruling supports what AI developers have been arguing for a while:
“Training AI on legally obtained text is transformative - it’s like human learning by reading.”
Meanwhile, Judge Alsup was direct in his decision:
“Fair use is not a loophole for piracy.”
What That Means (In Human Words)
There are two Anthropics in this case:
-
The one buying and scanning books to train Claude? ✅ Legal.
-
The one downloading troves of pirated material off the internet? ⚠️ Heading to trial.
This ruling is not a get-out-of-jail-free card for all AI training. It’s a signal that how you get your data matters - not just what you do with it.
Let’s Connect the Dots
AI lawsuits were inevitable. Until there are general governance frameworks and clear global laws around what is fair use and what is copyright infringement, courts will be busy addressing these cases one by one.
📖 Kadrey v. Meta
Yes, Anthropic isn’t the only one in court for training an AI model on books and whether it counts as fair use.
-
What was in question: 13 copyrighted books, some from shadow libraries
-
What the court took under consideration: Looking for proof of market harm
-
Outcome: As no proof of market harm was identified, it was considered fair use
The AI Act (EU)
The AI Act is considered, in our opinion, to be the first attempt to a designed process and a starting point guideline to achieving answers for legal issues when it comes to AI and data - answering the question of what is fair use and what is infringement.
1. It’s Allowed - If You Follow the Rules
The AI Act doesn’t ban training on copyrighted books. It builds on the EU's Text and Data Mining (TDM) exception:
“You can use copyrighted content for AI training unless the rights-holder has opted out.”
So:
-
If no opt-out? ✅ You’re clear.
-
If rights-holders opt out? ❌ You must not use their content.
2. You Must Disclose It
Under Article 28b (Amendment 399): AI providers must publicly disclose a summary of the copyrighted data used for training. This includes:
-
Books
-
Music
-
Videos
-
News articles
3. Why It Matters
Unlike the U.S., where courts interpret fair use case-by-case, the EU AI Act sets a clearer path:
-
Opt-outs must be respected
-
Disclosures are mandatory
-
Transparency becomes a legal expectation
This raises the bar for anyone training LLMs in or for the European market.
Bottom Line
This is a split ruling with serious consequences:
-
✅ Fair use win: Lawfully purchased and transformed data = legal (in the U.S.)
-
❌ Trouble ahead: Anthropic still faces a December trial over pirated books
🧠 From Thoughts to Prompt
Copyright is one of the biggest unresolved questions in the AI world.
And even though we don’t believe you have anything to do with the valuation side :)
-we do want to help you stay informed.
So we’ve written a prompt you can copy and paste into your LLM bot.
It will give you the latest updates based on your region
and help you understand what’s legal, what’s grey, and what’s a complete no-no.
Prompt to Paste:
What are the latest updates on copyright laws and regulations around AI training data - especially when it comes to books, articles, or written content? Include major rulings, lawsuits, or new policies from the last 3–6 months.
Focus on [insert country or region here] - and explain whether AI companies are allowed to train models on copyrighted material, and what the legal risks or protections currently look like.
Stop the AI Cult - Using the power of perspective
Frozen Light Team Perspective
This is one of those AI headlines that looks like a win - until you read the footnotes.
Understand that the law system doesn’t know how to handle this issue or understand the claim or impact. This is not just directed at the legal system - it also applies to publishers, as seen in the case of Kadrey v. Meta where the publisher could not show market impact.
The court basically said: whatever applies to people goes for AI training.
"If you read a book you bought, you're fine. If you downloaded 7 million bootlegs, you're not."
That should be fair enough.
This is why we admire the AI Act - it takes a broader, bird's eye view, trying to tackle big questions and, above all, aiming for visibility and freedom of choice.
Maybe there's a possibility that training AI on your books and content could be a good thing - maybe it helps AI learn how to promote or recommend your work.
At this point, no one really knows. Time will tell. But transparency will help everyone have the ability to decide if it helped or harmed them.
