case-law

Kadrey v. Meta: Why Meta's Win Was Narrower Than the Headlines Said

By Nora Whitfield · August 8, 2026

Meta won its case. That's the headline, and it's true. What gets lost in the headline is that the judge who ruled for Meta also went out of his way to say, in the same opinion, that he expects most future AI companies to lose this exact argument. Kadrey v. Meta is a win with an asterisk attached by the judge himself, and the asterisk matters more than the win.

What the case was about

Sarah Silverman, along with a group of other authors, sued Meta, alleging the company used pirated copies of their books, obtained through shadow libraries, to train its Llama models. The case was consolidated with a related action, Silverman v. Meta, and the core question was familiar territory by now: is training a large language model on copyrighted text a fair use.

Note the defendant here, because it's easy to confuse with a different case involving a plaintiff of the same name: this is Silverman and other authors against Meta, over book piracy in Llama's training data. It is a completely separate lawsuit from Authors Guild v. OpenAI, which involves different named plaintiffs (Tremblay, Chabon, and a separate Silverman claim) suing OpenAI specifically over ChatGPT's training data. Different defendant, different case, both currently active, easy to conflate if you're skimming headlines.

Judge Chhabria's ruling

On June 25, 2025, Judge Vince Chhabria in the Northern District of California granted Meta partial summary judgment on the fair-use question. The core holding: training an LLM on copyrighted text, without the model copying or outputting recognizable expression from that text, is a fair use. The transformation from "book text" to "statistical patterns a model learns from" was enough, in the court's view, to satisfy fair use's requirement of a new, non-substitutive purpose.

But Chhabria was explicit that this outcome turned on a specific failure in how the plaintiffs argued their case, not on a general principle that AI training is lawful. The plaintiffs' theory of harm centered on the idea that Llama, once trained on their books, could produce competing content that displaces the market for the original works, a theory known as market dilution. Chhabria found the plaintiffs simply hadn't put forward the evidence to support it. They hadn't shown, with anything concrete, that Llama actually was diluting the market for their specific books. Summary judgment on fair use, he wrote, was appropriate on the record in front of him, specifically because that record was thin on market harm evidence.

The part that matters more than the ruling

Here's where the case gets genuinely interesting, and where a lot of coverage undersold what happened. In the same opinion, in dicta (statements not strictly necessary to the ruling itself, but reflecting the judge's reasoning), Chhabria stated that he expects AI training on copyrighted work will often turn out to be illegal, once plaintiffs bring the right evidence of market dilution. He wasn't hedging out of caution. He was signaling, directly, that Meta won this round because of a gap in the plaintiffs' evidentiary presentation, not because the underlying legal theory of market harm is weak. Future plaintiffs who show up with real economic analysis, expert testimony on displaced sales, actual market data, could win on the exact same legal theory that just lost here.

That's an unusual thing for a judge to say in an opinion that otherwise rules against the plaintiffs, and it's why serious observers of AI copyright law don't read Kadrey as "training is legal." They read it as "this specific case, with this specific evidentiary record, lost, and here's exactly what a stronger case would need."

What's still alive

The ruling didn't end the litigation. It resolved the fair-use question on training specifically, for these plaintiffs, on this record. Distribution-based claims, allegations tied to how Meta obtained and redistributed the pirated material in the first place, rather than the training use downstream, were allowed to proceed. That's the same basic structural split you see in Bartz v. Anthropic: courts treating "how you got the material" as a separate question from "what you did with it once you had it," and being far more skeptical of the acquisition step than the training step.

What this means for AI authors

If you're trying to understand whether AI training on copyrighted work is settled law, Kadrey is direct evidence that it isn't, despite Meta technically winning. A single ruling that turns on a specific evidentiary gap isn't the same as a rule that would survive a better-argued case, and the judge who wrote this one told you that himself.

For your own practice, the actionable lesson is about evidence, the same theme running through nearly every case in this pillar. If you're a creator wondering whether an AI company's use of your work is causing you real, provable harm, "I feel like it's hurting my sales" isn't a legal argument. Concrete, documented market impact is. That's a hard thing to build after the fact. It's much easier to establish if you've been tracking your own sales, licensing, and market position over time as part of a broader record of your creative work and its commercial life.

The other lesson cuts toward AI users specifically: don't treat "a court said AI training can be fair use" as a blanket legal green light. This ruling was narrow, fact-specific, and explicitly flagged by the judge as unlikely to hold in a better-argued future case. The law here is actively moving, not settled, and building your own practice around records of what you actually made and how, rather than around a single favorable headline, is the more durable strategy regardless of which way individual training-fair-use rulings ultimately land.

Copyrightable is not a law firm and doesn't provide legal advice. Distribution-based claims in this litigation remain active. Confirm current status at official court dockets before relying on this summary.