Picture a conference room in a data center adjacent office, three engineers and one very tired lawyer staring at a spreadsheet of training corpus sources. Column A: dataset name. Column B: number of tokens. Column C, added last week and still mostly empty: provenance. Somebody asks whether the books shard came from a licensed vendor or a scrape. Nobody knows. The lawyer closes her laptop.
That moment is where the American copyright fight over AI training has actually landed. Not on whether models may learn from books, but on whether anyone can account for how the books arrived.
What the courts said
In the first two district court decisions on generative model training, Judges Alsup and Chhabria concluded that Anthropic’s and Meta’s training on copyrighted works fell within fair use. In Bartz v. Anthropic PBC, a group of authors sued over the use of their books; the court granted Anthropic summary judgment on fair use as to the training itself. Kadrey v. Meta Platforms, Inc. arrived at a parallel conclusion out of the same district. A federal court has since described this kind of training as “quintessentially” transformative fair use.
The word doing the work there is transformative. Fair use analysis asks, among other things, whether a new use adds something different in purpose or character rather than substituting for the original. Courts looked at training and saw a process that produces a statistical artifact, not a competing edition of the book.
The critical qualification: this reasoning covers legally acquired materials. Claims involving pirated copies were not swept up in the same holding. In the Anthropic matter, those copies were set to proceed toward trial to be decided separately. Acquisition and use are being treated as two distinct legal questions, and only one of them has gone the industry’s way.
Why this reads differently from an architecture perspective
I spend most of my time thinking about how agent systems store, retrieve, and recombine knowledge, and the legal reasoning here maps onto a technical distinction I find genuinely useful.
Training a large model on text produces parameters. Those parameters encode statistical regularities across an enormous corpus, and for the overwhelming majority of works in that corpus, no individual passage is recoverable from the weights. That is the argument for transformation, and it holds up reasonably well at the level of the base model.
Agent architectures complicate it. Modern agentic systems rarely rely on weights alone. They wrap retrieval layers, vector stores, document caches, and long context windows around the model. A retrieval-augmented pipeline that pulls a chapter into context and reasons over it is doing something structurally different from gradient descent over a token stream. The chapter is present, verbatim, at inference time. Nothing about it is statistical.
The decisions so far addressed training. They did not hand the industry a general license for every system that touches a copyrighted file. If you build agents, that gap is where your risk lives.
What this means for people actually shipping systems
A few implications I would treat as operationally solid, even while appeals and further litigation continue:
- Provenance is now an engineering requirement. If the legality of your training run depends on lawful acquisition, then dataset lineage stops being a documentation nicety and becomes part of your compliance surface. Track it the way you track model versions.
- Shadow libraries are a separate exposure. The distinction courts drew between training and pirated acquisition means a defensible use can sit on top of an indefensible source. Cleaning the source is cheaper than litigating the stack.
- Retrieval layers deserve their own analysis. Do not assume a training-focused holding covers your vector database, your document ingestion service, or your context assembly logic. Those systems reproduce text rather than abstract it.
- Output filtering matters more than it did. Transformation arguments weaken when a system emits substantial portions of a protected work. That is a testable property, so test it.
The unresolved part
Two district court opinions from one district are a beginning, not a settled body of law. Legal commentary, including analysis in the Houston Law Review, has framed these as first decisions rather than final ones, and recent rulings have raised further questions. Cross-border discussion of fair use adds another layer, since the doctrine is distinctly American and other jurisdictions apply different frameworks to the same technical process.
What I take from this is narrower and more actionable than the headlines suggest. The learning is defensible. The library card is not optional. And the further your architecture drifts from pure training toward systems that hold, retrieve, and echo source text, the less these rulings protect you.
Build accordingly. The provenance column in that spreadsheet is not going to fill itself in.
🕒 Published: