DOJ Advocates for Fair Use in AI Training to Protect National Security

On September 1, the U.S. Department of Justice (DOJ) submitted a Statement of Interest to the U.S. District Court for the Southern District of New York, advocating that the use of copyrighted written works for training large language models (LLMs) should be considered fair use. The DOJ contends that a ruling against this practice could distort copyright law, stifle innovation, and undermine U.S. competitiveness and national security. This statement was made in the context of ongoing copyright litigation involving OpenAI, particularly addressing claims from The New York Times, while also noting its relevance to cases involving authors and publishers.

The DOJ's argument highlights the distinct nature of generative AI training, which it asserts differs from merely collecting works and generating outputs. It emphasizes the need for courts to evaluate each alleged use of copyrighted material on a case-by-case basis. The government warns that imposing licensing requirements for model training could create barriers that only major tech companies could navigate, potentially consolidating the market for LLMs and favoring established media organizations with extensive archives.

In its brief, the DOJ focused on the first and fourth statutory fair-use factors, which pertain to the purpose and character of the use and its impact on the market for the copyrighted work. The government argues that the training process is highly transformative, as it does not utilize articles to entertain or inform in the way the original authors intended. Instead, it converts data into numerical representations, learning statistical relationships that enable the model to perform various tasks, including editing, translation, and generating new content.

Citing precedents such as Google v. Oracle and the Second Circuit’s Authors Guild v. Google ruling, the DOJ asserts that copying can be deemed fair when it facilitates new technological functions, even if entire works are reproduced during the training phase. The DOJ also notes that OpenAI's commercial intent should not weigh heavily against the fair use argument, as commercial uses can qualify as fair, especially when the use is transformative.

While acknowledging that certain outputs from LLMs could raise different copyright concerns, the DOJ maintains that these outputs should not dictate the assessment of whether the training use itself is transformative. It argues that a few anomalous outputs should not justify broad restrictions on model training or expose all LLM applications to significant liability.

Regarding market harm, the DOJ's filing emphasizes that copyright law recognizes harm from significant competition, not merely competition in general. It argues that training does not make protected expressions available to the public and thus does not replace or harm the market for the original works in a legally relevant sense.

The DOJ also criticized a 2025 ruling in Kadrey v. Meta Platforms, where a judge suggested that copying works for training LLMs would typically be considered infringing. The DOJ contended that this ruling improperly conflated training and output into a singular use and misinterpreted generalized competition as market harm.

To illustrate its position, the DOJ referenced Joan Didion's practice of typing Ernest Hemingway's stories to study his writing style, arguing that under the Kadrey reasoning, Didion could owe Hemingway whenever her writing competed in the literary market, which contradicts the principle that copyright protects expression rather than ideas or methods.

The DOJ concluded that imposing broad copyright liability that would generally prohibit AI model training without licensing would be both problematic and legally incorrect. It highlighted the public benefits of LLMs, which can assist various professionals, including writers and national-security officials, and help independent publishers compete against established entities. The filing noted that writers at The New York Times reportedly utilize AI to conceptualize and edit their articles, further underscoring the potential advantages of AI in the creative process.