The DOJ just told a court how to decide whether AI training is fair use

The fair use test that resolves every US copyright fair-use question has four factors, and a federal court in Manhattan is about to apply all four to a question that didn't exist when the test was codified: whether training a large language model on The New York Times' archive is infringement or fair use. The Trump administration filed its position this week, and the position is that the test's purpose-and-character factor, backed by the Constitution's "Progress of Science and useful Arts" clause, should resolve the case in OpenAI's favor.

The four factors are statutory, written into Section 107 of the Copyright Act. They ask, in order: what was the use for (commercial or educational, transformative or substitutive), what was the original work (factual or creative, published or unpublished), how much was taken (qualitatively and quantitatively), and what happens to the market for the original (does the new use substitute for it, or leave it alone). Judges weigh them together, not as a checklist.

The government is now telling the court how to weigh them. US attorneys argue in a statement of interest filed in The New York Times Co. v. Microsoft Corp. that The Times is trying to "narrow fair-use doctrine to exclude the training of OpenAl's large language models (LLMs)". LLM training, in the government's view, is "highly transformative" because the model does not republish the underlying articles, it learns patterns from them. That is the purpose-and-character factor doing most of the work in the filing.

Then the government leans on constitutional text. "Constraining LLM development under a misunderstanding of fair use doctrine would thwart such creative and scientific progress while hindering American prosperity and economic mobility," the filing argues, echoing the Constitution's instruction that copyright exists "to promote the Progress of Science and useful Arts". The argument is that if the four factors can plausibly tilt either way, the constitutional purpose of copyright itself should tip them toward OpenAI.

That move has a real precedent footprint. The judge overseeing this case, Sidney H. Stein, has already issued a memorandum explaining the April 2025 order on the parties' motions, and the case is widely tracked as the bellwether for AI-training litigation. The government's filing does not bind the judge. Statements of interest are policy signals that courts weigh, not directives they must follow.

The market factor is where the government's framing gets tested. The Times sued OpenAI and Microsoft in December 2023, seeking billions of dollars in damages for alleged training on its articles without permission. If LLMs trained on Times reporting can spit out answers that substitute for a Times subscription, that cuts against fair use. If they only learn from the reporting and refer users back to it, the market harm is harder to show. Several publishers have already settled the question by signing licensing deals with OpenAI, including the Associated Press, Axel Springer, and Vox Media, which suggests at least some publishers see a market for LLM access rather than a threat from it.

The most recent test of the market factor came from a separate case this year. In Bartz v. Anthropic, a federal judge ruled in 2025 that Anthropic could lawfully train on books it had purchased, but remained liable for training on pirated copies; the authors' class settled for about $1.5 billion. The takeaway is that the doctrine splits on acquisition: legitimate purchase is fair use, freebooting is not. That distinction does not resolve the Times case, because OpenAI's training data sourcing is part of the dispute, but it does set a hard outer edge on how far a "training is fair use" theory can stretch.

The Times is not the only publisher in court. The Center for Investigative Reporting, the Chicago Tribune, and the New York Daily News have filed parallel suits. A ruling against OpenAI would give those plaintiffs a ready template. A ruling for OpenAI, especially one leaning on the constitutional framing the government is now endorsing, would push more publishers toward the licensing table.

The political backdrop is hard to ignore. President Trump is separately suing The New York Times for defamation, and his administration has used statements of interest in private litigation as a recurring policy lever, with one official calling the tactic "incredibly" successful. The institutional fair-use position and the personal animus toward the Times are distinct matters, and the filing is the former. Treating them as the same would collapse a doctrinal question into a political one.

The court is being asked to apply a statutory test, not to ratify a constitutional theory. The four factors are written into Section 107, the Progress-of-Science clause is supplementary context, and the Bartz v. Anthropic precedent already draws a line between purchased training data and pirated training data. The judge's next written order in the Times case will show whether the four factors, on their own, are enough to settle the question before trial.