Washington picked a side. It picked OpenAI.
The Trump administration filed a 20-page brief in the New York Times’ lawsuit against OpenAI, arguing that training large language models on copyrighted material qualifies as fair use. The language is not hedged. “The United States has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law,” the filing says, pointing to scientific advancement and national security as the reasons. Making it harder for models to train on copyrighted content would be “inconsistent with basic copyright law principles” and would thwart creative and scientific progress, according to the brief.
I review tools for a living. I poke at APIs, run the same prompts through eight different wrappers, and tell you which ones fall apart under load. So my first instinct with a story like this is to ask the boring question: does this change what I’d recommend on Monday morning?
Why a court filing lands in your toolkit
Here is the thing I keep coming back to. Every AI tool you use sits on top of a model, and every model sits on top of a pile of training data. When you evaluate a coding assistant or a research agent or a writing tool, you are implicitly betting that the thing underneath it will still exist in eighteen months, priced roughly the way it is priced today.
That bet has a legal component whether you think about it or not. If courts had started treating unlicensed training as infringement, the cost structure of every model provider would shift. Licensing deals with publishers are not cheap, and those costs do not vanish into thin air. They land in your API bill, or they land in a smaller model trained on less data, or they land in a provider quietly shutting down the tier you built your workflow on.
A brief from the federal government does not settle the case. Courts are free to disagree with the executive branch, and the New York Times has its own arguments to make. But it is a meaningful signal about which way institutional weight is leaning, and that matters for anyone planning a stack more than a quarter out.
What I’m not going to pretend
I am not going to pretend I have no opinion on the underlying question. I write words for a living too. The idea that a model can absorb a decade of somebody’s reporting and then answer questions that used to send readers to that reporting is not a neutral technical detail. It is a real transfer of value, and the people on the losing end of it are not being unreasonable when they complain.
I also think the government’s national security framing deserves a raised eyebrow. It is the argument you reach for when you want a question decided quickly and in your favor. It may well be correct. It is also very convenient.
What I try to avoid in reviews is pretending that a tool’s ethics and a tool’s utility are the same axis. They are not. A model can be genuinely useful and built on a foundation you find uncomfortable. Both things can be true at once, and the honest move is to say so rather than picking whichever half fits the narrative.
The practical read
If you are making decisions about tooling right now, a few things follow from this.
- Legal risk is not the near-term threat to your stack. The bigger risks remain the ordinary ones: providers deprecating models, pricing changes, rate limits, and vendors who wrap an API and add nothing.
- Don’t build single-provider lock-in on principle. This filing is a signal, not a ruling. Tools that let you swap the underlying model are worth a small amount of extra friction.
- Provenance claims are now a differentiator, not a default. Some vendors market fully licensed or public-domain training data. If that matters to your organization, verify it rather than trusting the landing page. In my experience the claims are often looser than the copy suggests.
- Watch what publishers do next, not just what they say. Paywalls tightening and bot blocking will affect the freshness of the tools you use more directly than any court opinion will.
Where this leaves us
My honest take is that this story is more important for what it reveals than for what it decides. The federal government is treating the ability to train models on existing text as a national interest. That is a position with consequences well beyond one lawsuit, and it suggests the direction of travel for policy over the next few years.
For the tools on your desk today, nothing changes this week. The assistant that was good yesterday is still good. The wrapper charging you forty dollars a month for a system prompt is still not worth it. Keep evaluating on the same grounds you always did: does it do the work, does it fail gracefully, and can you get your data out when you want to leave.
Just know that the ground your stack stands on is being argued over in a courtroom, and one very large party has now shown its hand.
đź•’ Published: