The Justice Department just threw its weight behind OpenAI and Microsoft in the copyright fight that could define how AI models get trained for the next decade. In a filing tied to the consolidated lawsuit brought by The New York Times and other rights holders, the DOJ argues that training large language models on copyrighted text doesn’t amount to infringement. The reasoning hinges on a distinction the department considers fundamental: copying works during training is one thing, but those copies never surface publicly, and what the model actually produces “often if not always lack substantial similarity” to the originals.
The filing leans on an unexpected literary reference — Joan Didion. As a teenager, Didion copied Hemingway’s sentences by hand to understand how they worked. Under the logic of the Kadrey ruling, the DOJ argues, Didion could have faced liability every time she published because her learning process and her writing would be treated as a single use. The department calls that outcome unthinkable. It’s a clever analogy, but it sidesteps the obvious problem: a single author studying craft is not the same as a multibillion-dollar company ingesting millions of articles to build products that compete directly with the publisher it scraped.
The Trump administration’s position here is notable. The DOJ also takes direct aim at the US Copyright Office report that rejected blanket fair use for AI training, calling former Copyright Register Shira Perlmutter’s assessment legally non-binding and accusing the report of ignoring case law on market harm. Perlmutter was fired shortly after that report came out — Democrat Joe Morelle claims she was pushed out for refusing to legitimize AI training on copyrighted works, a stance favored by Trump ally Elon Musk. The DOJ’s footnote confirming Perlmutter is challenging her dismissal adds an awkward layer to an already politicized dispute.
What’s genuinely at stake is whether courts will treat training and output as legally distinct acts. The DOJ says yes, and that liability for training alone would stifle the very creativity copyright law exists to protect. Critics counter that this framing ignores scale entirely — the Copyright Office’s point that AI works with perfect copies at speeds no human could match. If the court adopts the DOJ’s reasoning, the NYT’s demand for billions in damages and the destruction of models trained on its articles collapses. If not, every AI lab in the country suddenly needs a licensing strategy. Either way, this case was already a bellwether. Now it has the full weight of the federal government on one side.
