Two district court decisions out of the Northern District of California did more to shape how we build language models than any framework release last year. Kadrey v. Meta Platforms and Bartz v. Anthropic both landed on the same core conclusion: training a model on copyrighted books can be fair use. One court called it “quintessentially” transformative. If you build bots for a living, that sentence is worth more than a hundred benchmark charts.
But the rulings came with a catch that a lot of the coverage glossed over, and it’s the part that actually touches your pipeline.
What the courts said yes to
The question in front of Judges Alsup and Chhabria was narrow. Not “is AI good,” not “should authors get paid,” but whether the specific act of feeding copyrighted text into a model during training counts as fair use under U.S. copyright law. Both judges weighed the fair use factors and came down on the side of transformation. A model that reads a novel and adjusts weights is not producing a substitute for that novel. It’s doing something categorically different with the text.
In Anthropic’s case, the court granted summary judgment on fair use for the training itself. Authors who sued over their books being used to train the models lost that piece of the argument.
If you’ve ever tried to explain to a client why your fine-tuned support bot isn’t “copying” the manuals it learned from, you now have a district court doing the explaining for you.
What the courts did not say yes to
Here’s where the ruling gets teeth. Fair use covered the training. It did not cover how the books got onto the disk in the first place.
The pirated copies in the Anthropic case were carved out and sent toward trial. The legal reasoning is straightforward once you see it: transformative use is a defense for what you do with a work, not a blanket pardon for how you obtained it. Buy the book, scan the book, train on the book — the court is largely with you. Pull a torrent of a shadow library and train on that, and you have a separate problem that fair use doesn’t solve.
For builders, that’s the whole ballgame. The legality of your model has less to do with your architecture and more to do with your procurement receipts.
What this changes in a real pipeline
I’ve been reworking how I handle data ingestion for the projects I run, and the shift is less dramatic than you’d expect. Mostly it’s bookkeeping that should have existed anyway:
- Log provenance per source, not per dataset. “Scraped from the web in 2025” is not provenance. A URL, a fetch timestamp, a license string, and a copy of the terms you agreed to is provenance.
- Treat aggregated datasets as untrusted input. If you pull a large text corpus off a model hub, you inherit whatever is inside it, including anything acquired badly. Read the dataset card. If it can’t tell you where the text came from, you can’t tell your legal team either.
- Separate acquisition from training in your codebase. Two stages, two logs, two audit trails. When someone asks how a document entered your system, you shouldn’t have to reverse-engineer a training script to answer.
- Keep purchase records with the data. Boring, but it’s the artifact that maps directly onto what the courts cared about.
None of that requires new tooling. A manifest file and some discipline covers most of it.
Why I’m not calling this settled
These are district court decisions. They’re persuasive, they’re detailed, and they’re the first real judicial reasoning we’ve had on generative AI training — the Houston Law Review piece walks through both opinions as exactly that, the first two. But district rulings are not national law. Appeals happen. Other circuits can reason differently. The pirated-copy claims are still headed for trial, and how that goes will shape the risk calculus for anyone who built on murky data.
There’s also a gap between what’s legal and what’s defensible to the people you work with. Clients, publishers, and open source communities have their own standards, and “a judge in California said it was fine” is a thin answer when a writer asks why their work is in your training set.
The practical takeaway
The transformation argument held up. That’s genuinely good news for small builders who don’t have a licensing department, because it means the technique itself isn’t the liability. Your exposure lives in your sourcing.
So the least glamorous part of your stack just became the most legally significant one. Go look at where your text comes from. If you can’t answer that in a sentence per source, that’s your next sprint.
🕒 Published: