The New York Times says OpenAI copied its journalism without asking. The United States government says a court should reject the argument that doing so breaks copyright law at all.
Both of those things are now on the record in the same lawsuit, which is a strange place for a country to end up. I’m Maya, and I spend most of my time here explaining AI agents to people who did not sign up for a law degree. So let’s walk through what actually happened, in plain language, and what it means for the tools you use.
What happened
The New York Times sued OpenAI over the use of its copyrighted articles to train large language models. That part has been going for a while. The new development is that the Trump administration filed a 20-page brief in the case, and it filed on OpenAI’s side.
The government’s position, in its own words, is that “the United States has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law.” Two reasons get named: scientific advancement and national security. The brief also argues that making it harder for large language models to train on copyrighted material would be “inconsistent with basic copyright law principles” and would hold back creative and scientific progress. TechCrunch’s Amanda Silberling covered the filing.
So the federal government is not just declining to referee. It’s arguing that training on copyrighted text is fair use.
Why training on text is the whole ballgame
If you’ve read anything on this site before, you know my favorite way to describe a language model: it’s a machine that learned patterns in language by reading an enormous pile of text. Not a database that stores articles, and not a search engine that looks things up. A system that adjusted millions of internal settings based on what it read, until it got good at predicting what words come next.
That distinction is exactly what the legal fight is about. Publishers look at the pile of text and see their work being copied and used commercially without a license. AI companies look at the same pile and say the reading itself is transformative, and that the output is a new thing, not a reproduction.
Fair use is the legal doctrine that decides which framing wins. It’s the same principle that lets a critic quote a novel or a researcher analyze a film. It’s flexible by design, which is why nobody can tell you with certainty how this comes out.
Why does this matter for AI agents specifically? Because agents are built on top of these models. The agent that drafts your emails, sorts your invoices, or researches a topic for you inherits everything the underlying model learned. If training data gets restricted, the models get narrower, and the agents built on them get less useful. That’s the practical chain of consequences.
What the government’s argument really signals
Here’s what I find notable. The brief doesn’t lean primarily on a creative-industry argument. It leans on national interest. Scientific progress. National security.
That’s a framing choice, and it tells you how the administration is thinking about AI. Not as a media dispute between a newspaper and a startup, but as infrastructure. The unstated logic goes something like this: if American models can’t train on the best available text, someone else’s models will, and the United States loses ground.
You can find that argument persuasive or convenient depending on where you sit. If you’re a working writer, a court hearing from the federal government that your archive is fair game for training is not a comforting development. If you build with these tools, it lowers a very large cloud of uncertainty that’s been hanging over the whole space.
What this doesn’t settle
A brief is an argument, not a ruling. The government is a powerful voice in a courtroom, but it isn’t the judge. The case is still the case.
A few things to keep straight as this moves along:
- Training and output are different questions. Whether a model can learn from an article is separate from whether it can reproduce one. Those get argued on different terms.
- Fair use is decided case by case. A finding in one lawsuit doesn’t automatically settle every other claim against every other model.
- Licensing deals exist alongside the litigation. Legal permission and commercial partnership aren’t mutually exclusive, and plenty of publishers have pursued both paths.
My honest read
The most useful thing you can take from this is that the question of what AI is allowed to read is now openly a policy question, not just a legal one. That’s a shift in where the argument lives.
For anyone using AI agents day to day, nothing changes tomorrow. Your tools still work the same way. But the rules being written now determine what those tools are able to know in five years, and who gets paid along the way. That’s worth following, even if the filings are dry reading.
I’ll keep translating as it develops.
🕒 Published: