Judge William Alsup called training AI on copyrighted books “quintessentially” transformative. That single word, coming from a federal judge in the Northern District of California, did more for AI builders than a year of think-pieces. And then the same judge ordered Anthropic to pay a $1.5 billion settlement to a group of writers.
Both of those things are true at once. If that seems contradictory, it isn’t. It’s the most important distinction anyone building with models needs to understand right now.
The split that matters
In Bartz v. Anthropic, the court separated two questions that most people mash together. Question one: is it legal to train a model on copyrighted books? The court’s answer leaned yes, on transformative fair use grounds. Question two: is it legal to assemble a central library of pirated copies to do that training? On that, the court denied summary judgment for Anthropic. That’s the part that turned into a $1.5 billion line item.
So the legal risk didn’t live in the training. It lived in the acquisition. Kadrey v. Meta Platforms came out of the same district and pointed in a similar direction. Training can be fair use. Pirating may still be infringement.
As someone who spends most of my week wiring together retrieval pipelines and fine-tuning jobs, this reframing hit me harder than any of the doomsday coverage did. I had been thinking about my data the way I think about a dependency: does it work, is it clean, does it parse. I was not thinking about it the way a lawyer thinks about it: where did this come from, and can I prove it.
What this changes about how I build
Most tutorials on fine-tuning start at the same place. Grab a dataset. Load it. Tokenize. Train. Nobody tells you to ask who owns the text. The scraped corpus in that HuggingFace repo with 40,000 downloads has a provenance story, and you inherit it the moment you use it.
Here’s the mental shift I’d suggest for anyone shipping bots that touch external text:
- Provenance is a build artifact. Track where every corpus came from with the same seriousness you track package versions. A manifest file listing source, license, acquisition date, and method costs you twenty minutes and answers the only question that matters if anyone ever asks.
- “Publicly available” is not a license. A book being findable online says nothing about whether the copy you found was legally distributed. Shadow libraries are publicly available. They are also the thing that generated a ten-figure settlement.
- Buying the book is a real strategy. It sounds almost quaint, but legal acquisition was the hinge in these rulings. Purchased copies sit on much firmer ground than downloaded ones.
- Transformation is about output, not intent. The court’s reasoning centered on the use being transformative. If your system reproduces long passages verbatim, you’re closer to distribution than to transformation, whatever your training pipeline looked like.
The RAG wrinkle nobody talks about
Retrieval-augmented generation deserves its own paragraph, because it sits in a genuinely different position than training. When you train, the copyrighted text is metabolized into weights. When you retrieve, you’re storing the actual text in a vector database and pulling chunks of it back out at query time.
These decisions were about training. I’d be careful reading them as blanket permission for retrieval systems that keep and surface source text. The transformative argument gets thinner the closer your architecture gets to “search engine that quotes books.” Chunk sizes, whether you return excerpts to the user, whether you cite and link to originals: these are now design decisions with legal weight, not just UX preferences.
Where things actually stand
Fair use analysis for AI continues to develop through 2026, with law firms publishing update after update as new decisions land. That pace tells you something. Nothing here is settled in the way engineers like things settled. Two district court decisions from one district are a signal, not a rulebook, and appeals reshape signals.
What I take from it is narrower and more useful than a verdict. The direction of travel favors builders who can document a clean supply chain for their data and disfavors those who can’t. That’s not a legal opinion, it’s an engineering observation about which teams sleep well.
If you’re building something today, the practical move is unglamorous. Write down where your data came from. Prefer sources you can point to a license or a receipt for. Design your outputs so they synthesize rather than reproduce. And check current rulings before you make a bet you can’t unwind, because this space is moving faster than any blog post can keep up with, including this one.
🕒 Published: