Twenty pages. That’s the length of the brief the Trump administration filed in support of OpenAI’s unlicensed use of copyrighted material in the New York Times lawsuit. For a question that touches nearly every large language model in production today, twenty pages is a remarkably compact intervention — and a telling one. Governments do not usually file in private copyright disputes over training data unless they believe something structural is at stake.
The government’s position, as reported, is direct: the United States has a strong interest in the court rejecting any argument that training LLMs on copyrighted texts violates copyright law. The reasoning leans on two pillars — scientific advancement and national security — plus the claim that making it harder for models to train on copyrighted content would be inconsistent with basic copyright law principles and would obstruct creative and scientific progress.
I want to look at this from the architecture side rather than the legal one, because the technical assumptions buried in that argument matter more than the rhetoric around it.
What the fair use claim assumes about how models work
Any fair use defense for training rests on an implicit technical claim: that a trained model is a transformation, not a container. That the weights encode statistical regularities about language, not retrievable copies of the inputs. If that claim holds cleanly, the legal case is relatively straightforward. If it holds only most of the time, things get complicated.
Anyone who has worked closely with these systems knows the picture is messier than either side’s talking points. Memorization is not binary. It scales with parameter count, with how many times a sequence appears in the corpus, and with how distinctive that sequence is. A wire story that got syndicated across two hundred outlets sits differently in a model’s weights than a single unique feature piece. The transformation argument is strongest for the broad statistical middle of a corpus and weakest at its tails — the rare, high-value, frequently-duplicated text that publishers care most about.
So the interesting question is not whether training is transformative in principle. It’s where a court decides to draw the line on the distribution, and whether anyone in the room has the tools to measure it.
Why this lands hardest on agent systems
For readers who spend their time on agent architecture, there’s a second-order effect worth tracking. Agents are not just models — they’re models plus retrieval, plus tool use, plus memory. The legal question in front of the court concerns pretraining. But the systems being built today mix pretrained weights with live document access in ways that blur the boundary the fair use argument depends on.
Consider the difference:
- A model that learned statistical patterns from a news archive years ago and produces prose in a similar register.
- An agent that fetches today’s article, holds it in context, summarizes it, and passes that summary to a downstream step.
- A memory system that caches those summaries and reuses them across sessions, building a persistent derivative of a copyrighted corpus.
Those are three different legal objects wearing the same product interface. A ruling that settles the first says little about the third. Agent builders who read a favorable outcome here as blanket permission will be building on a foundation that was never poured under them.
The national security framing deserves scrutiny
The security argument is the part I find least satisfying technically. The implication is that restricted training data means weaker domestic models, which means strategic disadvantage. That chain has a weak link in the middle. Frontier capability gains over the last few years have come at least as much from post-training methods, reasoning approaches, and inference-time compute as from raw corpus expansion. Data quality and curation have arguably mattered more than data volume for a while now.
If the actual constraint on model capability were access to newspaper archives, we would expect to see capability tracking corpus size much more tightly than it does. Treating training data access as a national security bottleneck flattens a more interesting engineering reality into a simpler political story.
What I’d watch next
The outcome here will shape a lot of downstream architectural decisions, and not always in obvious directions. A ruling that fully validates the fair use position removes pressure to build licensing infrastructure, provenance tracking, and data attribution into training pipelines. Those are technically useful capabilities regardless of the legal requirement — knowing what’s in your corpus makes debugging, auditing, and targeted unlearning far more tractable.
The uncomfortable possibility is that winning this argument decisively makes the systems we build less legible, not more. Legal pressure has been one of the few forces pushing labs toward documenting their data. Remove it and the incentive to invest in that plumbing largely disappears.
Whatever the court concludes, the technical questions underneath — how much a model retains, where, and whether we can measure it — remain unanswered. A favorable brief does not resolve them. It just delays the day we have to.
🕒 Published: