by Fulvio Sarzana di S. Ippolito, Attorney, Sarzana & Partners Law Firm
In a brief filed on 1 September in the New York litigation, the United States asks the court to rule that fair use covers copying protected texts to train an AI model, and that it does not infringe copyright. The filing attacks a California ruling and the report of the federal copyright office.
On 1 September 2026 the Department of Justice filed a Statement of Interest of the United States, the brief through which the federal government sets out its interest to a judge in a dispute between private parties. The case is In re: OpenAI, Inc. Copyright Infringement Litigation, consolidated proceeding 25-md-3143 (SHS) (OTW) before the District Court for the Southern District of New York.
The claims of The New York Times, The Intercept, Tribune Media, Ziff Davis, authors and book publishers have all been merged here. The caption of the filing, docket entry 316, states that it applies to all consolidated positions. US newspapers reported the news on 2 September.
The tool is 28 USC § 517, which lets the Department step into any case pending before a federal court to represent the interest of the United States. The brief notes that the provision sets no deadline and requires no leave of court. It is signed by Associate Attorney General Stanley E. Woodward Jr., Assistant Attorney General Brett Shumate and senior counsel Michael Weisbuch.
The request is blunt: the United States has «a strong interest in this Court rejecting any argument that training LLMs on copyrighted works infringes copyright».
Fair use: separating training from model output
The legal structure turns on a distinction the government repeats at every step. A model goes through more than one development phase, and each raises its own copyright questions. The Department deals only with training, that is, the copying of works to feed them to the model, and leaves aside what the model writes when a user queries it.
The first of the four fair use factors looks at the purpose of the use: whether the copier simply replaces the original work, or does something with it that serves a different end. For the Department, training clearly falls into the second case, because a New York Times article is copied not to be read, but to extract the statistical patterns of language, the vocabulary and the syntax.
The precedents cited are two California decisions from 2025, Bartz v. Anthropic and Kadrey v. Meta, and above all Google v. Oracle from 2021, where the Supreme Court held that the full copying of someone else's code was lawful because it served to build something the original did not do.
The fourth factor looks instead at harm to the market for the work. Here the brief argues that the copy used for training does not act as a substitute for the article and remains internal to the technical process: nothing of the source text reaches the public, because nobody reads the model, people read its answers. From there comes the sharpest passage of the filing.
The attack on Kadrey and the Copyright Office
The Department calls «deeply flawed» the part of Kadrey that built the theory of market dilution, that is, the indirect harm from the risk that the market gets flooded with works generated by the models. According to the government that theory confuses training with model output, ignores the substantial similarity requirement and finds no support in any precedent.
The example chosen to take it apart is literary. As a young woman, Joan Didion retyped Hemingway's stories to understand how the sentences worked. Under the logic of Kadrey, the Department writes, Didion would have owed Hemingway damages every time she published something.
Footnote 17 delivers the institutional blow. The government notes that the same theory had been endorsed by the Register of Copyrights, that is, by the head of the federal copyright office, in the third part of the report on artificial intelligence and copyright of May 2025. It adds that after Loper Bright a judge is not bound to defer to the reading of an administrative authority, and that the report has a «thin» reasoning anyway. It then recalls, in passing, that the Register has occurred over her own removal from office: the report had come out in pre-publication form the day before the administration dismissed her.
The industrial policy argument
The opening part of the brief is not legal. The United States invokes the executive order of 23 January 2025 on removing barriers to American leadership in artificial intelligence and the one of 2 June 2026, then national security and the risk that stricter rules would favor foreign adversaries.
An inverted antitrust argument follows. If training required licenses, the government argues, only the largest technology companies would have the capital to pay for them, and the fees would flow disproportionately to traditional publishers purely because of the volume of their archives. Licenses would become «in essence large subsidies for legacy media companies».
It is a reading that reverses the way Europe looks at the relationship between platforms and publishers. Here the related right in Article 15 of Directive 2019/790 exists precisely to correct a contractual imbalance in favor of publishers. In Washington, paying publishers is described as a barrier to entry.
Matt Topic, who represents The Intercept, replied that the administration's position amounts to «an unprecedented and uncompensated transfer of intellectual property rights from news organizations to technology companies», and noted that the Department argues without access to the discovery record.
Why American fair use matters for Europe too
In Europe fair use does not exist, and the question is framed the other way around. The starting point is the text and data mining exceptions in Articles 3 and 4 of Directive 2019/790, transposed in Italy into Articles 70-ter and 70-quater of the copyright law. For uses other than scientific research, the exception applies only if the rightholder has not reserved his rights.
So in Europe the debate is not about the nature of the use, it is about whether the refusal was expressed and whether it was respected. The rules on that refusal have tightened considerably over the past 2 years.
Article 53 of the Artificial Intelligence Act requires anyone placing a general-purpose model on the EU market to adopt a policy to comply with the reservation of rights, and to publish a summary of the content used for training following the template drawn up by the AI Office. The obligation applies even if training took place outside the Union. The code of practice of 10 July 2025 turns the principle into operational requirements: identify the reservation with up-to-date technologies, respect the robots.txt exclusion protocol in the RFC 9309 version, name a contact point for rightholder complaints.
In Italy, Law 132/2025 added Article 70-septies to the copyright law, which refers to the two exceptions and expressly applies them to model training. The text leaves open the question that matters, that is, how the rightholder's objection must be expressed so that a machine can recognize it.
The European Parliament has asked to go further. In an own-initiative resolution approved in March this year, with Axel Voss as rapporteur, it proposed a centralized register of reservations in machine-readable format, transparency obligations on training data and a system of collective licensing with remuneration for authors. The text binds no one, but it shows the direction the Commission is being pushed in.
The first European case, meanwhile, is already under way. In C-250/25 houses, Like Company v Google Ireland, a Hungarian court has asked the Court of Justice whether a chatbot answer that reproduces parts of articles is communication to the public for the purposes of the publishers' related right, whether training is reproduction, whether it falls within the Article 4 exception, and whether the provider is responsible for what the model returns at the user's request. These are the same phases the Department of Justice tried to keep separate.
Judge Sidney Stein will rule on fair use in a proceeding where one of the parties now has the federal executive at its side, with a filing that openly states its aim to shape the outcome of a dispute between private parties. That, more than the fair use doctrine itself, will be worth coming back to.
A first version of this article was published on 3 September 2026 on Fulvio Sarzana's blog at Il Sole 24 Ore (in Italian).
English 
Italian 