The Justice Department has filed a statement of interest in the consolidated copyright litigation against OpenAI, urging the Southern District of New York to hold that training a large language model on copyrighted written works is fair use. The filing landed on September 1 in In re OpenAI, Inc. Copyright Infringement Litigation, MDL No. 25-md-3143, before Judge Sidney H. Stein.
A statement of interest is a filing the federal government uses to state a position in a case it is not a party to. It binds nothing. No judge is obliged to give it more weight than any other brief on the docket. But it is, by several accounts, the first time the United States has taken a formal position in any of the copyright suits that authors, publishers, music labels and news organisations have brought against AI developers since 2023 — and that first is the story.
The government’s argument runs on two tracks. The first is conventional copyright doctrine: citing Google v. Oracle and the Second Circuit’s Authors Guild v. Google, the DOJ argues that copying can be fair when it enables a new technological function, even where entire works are reproduced at an intermediate step, and that OpenAI’s commercial purpose should carry diminishing weight as the use becomes more transformative.
The second track is not a copyright argument at all. The DOJ asserts a “profound national interest” in a dominant domestic AI industry and warns that a restrictive reading of fair use would suppress innovation and hand advantage to foreign competitors. Translated: licensing costs are a national security problem.
Why a non-binding brief still moves money
Almost every AI content-licensing deal signed in the last two years was priced against a probability — the odds that a court would eventually rule training is infringement and force retroactive payment. That probability is the entire basis of publisher leverage at the negotiating table. Nobody was paying because they had to; they were paying to buy certainty.
A statement of interest from the Justice Department does not change the law. It changes the number every general counsel writes in the margin next to “litigation risk.” Move that number a few points and the price of a licensing deal moves with it, in every negotiation currently open, immediately, regardless of what Judge Stein eventually rules.
The Times has said the administration is siding with a handful of trillion-dollar AI companies, and that AI firms should simply pay fairly for the content that makes their products work. That is the correct argument for a plaintiff to make. It is also, for now, an argument being made without the government in the room on its side.
Our take: The doctrinal half of this brief is unremarkable — the transformative-use argument is the one OpenAI has been making since 2023, and reasonable lawyers disagree about it. The national-security half is the part worth watching, because it imports a rationale that copyright law has no machinery to weigh. Fair use has four statutory factors and “does this help America beat China” is not among them. If that framing takes hold, the fight stops being about copyright and becomes an industrial-policy question decided in courtrooms that were never built for it.
What to watch
- Whether Judge Stein engages the national-security argument at all. A ruling that decides the case purely on the four statutory factors would quietly signal how much that framing is worth.
- Repricing in open licensing negotiations. Publishers with deals in flight now have a weaker hand than they did last week. Watch for terms getting shorter and cheaper rather than deals collapsing outright.
- Whether other agencies follow. A statement of interest is cheap to file. If this becomes a pattern across the related author and publisher cases, it stops looking like a one-off and starts looking like policy.
- Congress. Several readings of the filing push creators toward legislation rather than litigation. That is a much slower road, and a much more expensive one to lobby.
If you build on these models
Nothing about your exposure changed today. Indemnification clauses in your model-vendor contracts still do what they did last week, and you should still know whether yours covers output-side claims or only training-side ones. What changed is the probability distribution your vendor is pricing against — which means the next renewal conversation is being held on slightly different terms than the last one, whether or not anyone says so out loud.