U.S. Government Backs Fair Use for AI Training in OpenAI Copyright Litigation

Background

The U.S. government has filed a Statement of Interest of the United States of America under 28 U.S.C. § 517 in the Southern District of New York multidistrict litigation over OpenAI’s use of copyrighted works to train large language models. The filing appears to be the federal government’s first direct intervention in the wave of artificial intelligence (AI) training copyright cases, and it advises the court to treat model training (separate from use of AI to generate outputs) on copyrighted text standing alone as a highly transformative fair use under current law.

Previously, the U.S. Copyright Office released its prepublication “Copyright and Artificial Intelligence Part 3: Report on Generative AI Training” in May 2025 under Register of Copyrights Shira Perlmutter, following earlier report installments on digital replicas and copyrightability. As discussed in our May 2025 client alert, the Part 3 report treated the application of fair use to generative AI training as a fact-specific inquiry, with the outcome depending on such considerations as the works used, the source of the training material, the purpose of the model, its outputs, and the relevant licensing markets. Shortly thereafter, the Trump Administration attempted to remove Perlmutter, prompting litigation over control of the Copyright Office and the President’s authority to displace a Register who is appointed and removed through the Librarian of Congress. On June 30, 2026, the Supreme Court denied the government’s application to stay the D.C. Circuit’s order permitting Perlmutter to remain in office while the litigation proceeds, preserving the possibility that the Copyright Office will formally publish the final Part 3 report, as discussed in our prior alert on that ruling and its implications.

That backdrop, together with the Administration’s release of America’s AI Action Plan last year, makes clear the Executive Branch’s pro-training position before the court after the Copyright Office had issued a more cautious, fact-specific analysis, and it underscores the extent to which AI training has become both a copyright-law issue and an Administration-level policy priority.

The brief also contextualizes fair use in a broader national policy debate. The government ties AI development (or more specifically, development of large language models (LLMs)) to national security, economic competitiveness, scientific progress, and U.S. leadership in emerging technology. It warns that judicially imposed licensing requirements for training could slow domestic AI development, strengthen foreign competitors, and concentrate the market among companies able to absorb large-scale licensing costs.

For clients, the filing is important because it converts that policy debate into a litigation position. Rather than leaving the AI-training question to a broad factual spectrum, the government urges a practical rule: Courts should separate internal training from public outputs and should not treat the mere use of copyrighted text to train an LLM as infringement absent a distinct use that substitutes for protected expression.

The Administration’s Argument

The Administration’s argument is grounded in familiar fair-use doctrine, but the brief applies that doctrine in a way that would materially narrow training-stage copyright claims. Four points are likely to be most important as the OpenAI v. New York Times litigation moves toward summary judgment:

• Training vs. outputs. The government argues that fair use must be assessed use by use, so internal or intermediate copying to train a model should not be treated the same as public-facing outputs that may reproduce protected expression. That distinction matters because plaintiffs often point to allegedly infringing outputs, summaries, or memorized text to challenge training itself; the brief urges the court to address any output problem with output-focused remedies.
• LLM training is highly transformative. The brief characterizes training as using text to identify linguistic patterns, relationships, and predictive signals, not to substitute for the expressive purpose of the underlying works. In that view, the model-training use is different in kind from reading, displaying, selling, or republishing the copyrighted works themselves.
• Commercial use not dispositive. The Administration acknowledges that many LLM products are commercial but argues that commerciality should carry less weight where the challenged use is transformative and does not expose protected expression to the public. That framing is designed to keep the first factor focused on the purpose of training rather than on the business model of the AI developer.
• Harm requires substitution, not general competition. For the fourth factor, the brief rejects the theory that AI-generated works in the same genre or broader competition from AI-enabled content constitute copyright harm by themselves. Instead, the Administration argues that copyright-relevant market harm requires substantial similarity, substitution for protected expression, or impairment of a cognizable derivative market tied to the works at issue.
  
  

Further, the Administration does not foreclose voluntary licensing or future legislative solutions, but it argues that courts applying current law should not convert unresolved policy questions into a mandatory licensing regime for all training uses.

Departure from the Copyright Office’s Prepublication Part 3 Report

The most important context is that both documents arise from the same Administration-era policy environment but point in materially different directions. The Copyright Office’s May 2025 prepublication Part 3 report, issued under Register of Copyrights Shira Perlmutter shortly before the Trump Administration attempted to remove her, treated generative AI training as a fact-intensive fair-use question rather than a categorical answer.

The Part 3 report recognized that fair use may protect some training uses, particularly noncommercial research or analysis that does not enable reproduction of protected expression. But it also warned that copying expressive works from pirate sources to generate unrestricted content that competes in the marketplace, especially where licensing is reasonably available, is unlikely to qualify. The report treated the source of training materials, the purpose of the model, output safeguards, and emerging licensing markets as central facts in the fair-use analysis.

By contrast, the Administration’s brief asks the OpenAI court to isolate training as an internal, transformative use and to treat outputs, acquisition, and licensing as separate issues. That framing is a sea change from the Part 3 report because it shifts emphasis away from a spectrum of training scenarios and toward a broad rule that training on copyrighted text standing alone should not be treated as infringement.

The shift is especially notable on licensing. The Copyright Office described voluntary licensing markets as developing and suggested collective approaches could be considered if market gaps persist. The Administration’s brief, however, warns that court-imposed licensing obligations could hinder U.S. AI competitiveness and entrench large incumbents, making licensing primarily a legislative or a commercial issue rather than a judicial fair-use requirement.

The brief also departs sharply on market harm. Where the Part 3 report treated potential licensing markets and market-competing outputs as relevant to the fair-use inquiry, the Administration’s brief rejects broader “market dilution” theories and argues that AI-enabled competition does not create cognizable fourth-factor harm unless outputs are substantially similar to or substitute for protected expression.

Implications and Takeaways

The filing is not binding on the court, and it does not settle the unsettled fair-use questions now moving through multiple AI copyright cases. It is nevertheless likely to influence briefing in the OpenAI vs. New York Times litigation and related disputes because it gives defendants a clear federal policy statement to cite and gives copyright owners a preview of the arguments they will need to answer.

• For AI developers and enterprise users, the filing strengthens the argument that training-stage copying can be fair use where the model does not expose protected expression. But it should not be read as a blanket immunity: claims based on pirated-source allegations, output memorization, substantial similarity, or misleading use of copyrighted content may remain viable depending on the facts. There is also the possibility that the court rejects those arguments in the brief and rejects the fair-use defense for AI training entirely.
• In the near term, stakeholders should preserve evidence on training-data sources, model safeguards, licensing history, output controls, and any substantiated instances of reproduction or substitution. Those facts are likely to drive the next phase of fair-use litigation more than broad arguments about whether AI is beneficial or harmful in the abstract.