The Trump administration has put itself on OpenAI's side of the copyright fight over training data, arguing in a 20-page brief that the United States has an interest in a competitive AI industry that sets the global standard for how the technology is used. The document cites an executive order signed by President Donald Trump holding that the country must maintain its global leadership in AI, and warns that constraining the development of large language models through a mistaken reading of fair use would slow creative and scientific progress and damage economic development and social mobility in the United States. It is not a ruling. The case is before the US District Court for the Southern District of New York, and the brief's authors have no jurisdiction over it.
The dispute itself is unresolved law. The models behind ChatGPT, Claude and Gemini are trained on enormous volumes of published material — books, articles and other copyrighted work — and AI companies add that material to training sets without asking rightsholders. Many publishers, The New York Times among them, argue this breaks the law. Whether it does has no settled answer.
The argument runs through fair use, the doctrine that permits using someone else's work without permission in certain circumstances. In the training cases, courts have to decide whether what a model does with the source material is transformative enough to qualify.
So far, those decisions have mostly gone the technology companies' way. The apparent exception proves the point. Judge William Alsup ordered Anthropic to pay a group of writers $1.5 billion over works used to train its models — but the company was not penalised for the training. The penalty attached to how it obtained the books: Anthropic had pulled them from illegal shadow libraries. On the training question itself, Alsup compared a language model to a person who reads books because they want to become a writer, and concluded the model studies the works not to reproduce or displace them but to make something else.
That distinction is doing more work than the headline number suggests. The largest payout in this field to date was a procurement ruling, not a copyright-scope ruling, and it left the underlying permission intact. The rule it establishes is narrow and expensive: learn from whatever you like, but buy it.
The brief is best read as politics rather than law, and it does not hide it. The case for OpenAI is made almost entirely in terms of national competitiveness — standards, leadership, economic development — rather than in terms of what the copyright statute actually says. That is a legitimate thing for an executive branch to believe and an odd thing to file in a private infringement suit, where the question is not whether American AI should win but whether these specific books were used lawfully. Twenty pages is also a slight document to put against a question that has already produced a $1.5 billion judgement. This reads less like an attempt to win the argument than to be visibly present for it.
What the administration does not address is the part that has actually cost an AI company money. Alsup's ruling turned on acquisition, and nothing in a fair-use defence tells a lab where its books came from. If the administration's theory prevails, the constraint on frontier labs is not whether they may learn from copyrighted work — it is whether they can document paying for it. That is a compliance problem, and compliance problems favour incumbents: the companies that can afford licensing deals and clean data provenance are the ones already large enough to be sued.
A court in the Southern District of New York is not obliged to care what the executive branch thinks about fair use, and the brief carries no authority there. But judges read the room as well as the record, and an administration declaring that a ruling against AI training would harm American competitiveness has changed what a ruling against AI training costs the judge who writes it. That is the intervention, whatever the document says it is.