*by Fulvio Sarzana di S.Ippolito, Lawyer, Sarzana and Partners Law Firm. Sarzana Law Firm and Associates
In a brief filed on September 1st in the New York litigation, the United States asks the judge to declare that training a model on protected texts constitutes fair use and does not infringe copyright. The brief criticizes a California ruling and the Federal Copyright Office's report.
On September 1, 2026, the Department of Justice filed a Statement of Interest of the United States, the brief by which the federal government presents to a judge its interest in a case between private individuals. The procedure is In re: OpenAI, Inc. Copyright Infringement Litigation, consolidated proceeding 25-md-3143 (SHS) (OTW) in the District Court for the Southern District of New York.
The lawsuits filed by the New York Times, Intercept, Tribune Media, Ziff Davis, authors, and book publishers have been consolidated here. The title of the lawsuit, document 316 of the case, specifies that it applies to all the positions filed. The news was published in American newspapers on September 2nd.
The instrument is 28 USC § 517, which allows the Department to intervene in any case pending in federal court to represent the interests of the United States. The brief notes that the provision has no time limits and does not require judicial authorization. The signatories are Associate Attorney General Stanley E. Woodward Jr., Assistant Attorney General Brett Shumate, and Senior Counsel Michael Weisbuch.
The request is clear: the United States has "a strong interest in this Court rejecting any argument that training LLMs on protected texts violates copyright.".
Fair use: Separate training from model responses
The legal framework hinges on a distinction the government reiterates at every step. There are multiple development phases for a model, and each raises its own copyright issues. The Department only handles training, that is, copying the works to feed them into the model, and leaves out of its reasoning what the model writes when a user queries it.
The first of the four fair use criteria concerns the purpose of the use: whether the copier merely replaces the original work, or uses it for something that serves a different purpose. For the Department, the training clearly lies in the latter case, because the New York Times article was copied not to be read, but to extract statistical regularities of the language, vocabulary, and syntax.
The precedents recalled are two Californian decisions from 2025, Bartz v. Anthropic e Kadrey v. Meta, and above all Google v. Oracle of 2021, where the Supreme Court held that the complete copying of someone else's code was lawful because it served to construct something that the original did not do.
The fourth criterion instead addresses the damage to the work's market. Here, the brief argues that the training copy doesn't function as a substitute for the article and remains internal to the technical process: the public receives nothing of the source text, because no one reads the model; they read their responses. Hence the most difficult part of the process.
The attack on Kadrey and the Copyright Office
The Department defines the part of "deeply flawed" Kadrey who had built the theory of market dilution, that is, the indirect harm resulting from the possibility that the market will be flooded with works generated by models. According to the government, this theory confuses training with model responses, ignores the requirement of substantial similarity, and is unsupported by any precedent.
The example chosen to dismantle it is literary. Joan Didion, as a girl, typed out Hemingway's stories to understand how sentences worked. With the logic of Kadrey, the Department writes, Didion would have to compensate Hemingway every time she published something.
In footnote 17, the institutional blow arrives. The government notes that the same theory had been endorsed by the Register of Copyrights, i.e., 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 The judge isn't required to comply with an administrative authority's interpretation, and that report's reasoning is "weak" in any case. He then incidentally recalls that Register challenged her dismissal in court: the report had been released in a preliminary version the day before the administration fired her.
The industrial policy topic
The initial part of the brief is not legal. The United States invokes the executive order of January 23, 2025, on removing barriers to American leadership in artificial intelligence and the one of June 2, 2026, then national security and the risk that stricter regulations would benefit foreign adversaries.
A reversed antitrust argument follows. If training required licenses, the government argues, only the largest tech companies would have the capital to pay for them, and the fees would disproportionately accrue to traditional publishers based on the sheer volume of their archives. The licenses would "essentially become huge subsidies for legacy media companies.".
This is a reading that overturns the way we look at the relationship between platforms and publishers in Europe. Here, the related right of Article 15 of the Directive 2019/790 It was created precisely to correct a contractual imbalance that favors publishers. In Washington, payments to publishers are described as a barrier to entry.
Matt Topic, defending the Intercept, responded 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 is making its argument without access to the investigation's records.
Why American fair use concerns us too
In Europe, fair use does not exist, and the question is posed in reverse. The starting point is the exceptions for text and data mining in Articles 3 and 4 of Directive 2019/790, transposed in Italy into Articles 70-ter and 70-quater of the Copyright Act. For uses other than scientific research, the exception applies only if the owner has not reserved their rights.
So here we don't discuss the nature of the use, we discuss whether the refusal was expressed and whether it was respected. The rules on that refusal have become considerably more stringent in the last two years.
Article 53 of the regulation on artificial intelligence Anyone placing a general-purpose model on the European market must adopt a policy to respect the reservation of rights and publish a summary of the training content according to the template developed by the AI Office. This requirement applies even if the training took place outside the Union. The Code of Practice of July 10, 2025, translates the principle into operational requirements: identifying the reservation with up-to-date technologies, complying with the robots.txt exclusion protocol in RFC 9309, and indicating a contact point for owner complaints.
In Italy, Law 132/2025 added Article 70-septies to the Copyright Act, which recalls the two exceptions and expressly refers them to model training. The text leaves open the crucial issue of how the owner's objection must be expressed for a machine to recognize it.
The European Parliament has called for further action. In an own-initiative resolution adopted in March this year, with Axel Voss as rapporteur, it proposed a centralized machine-readable register of reserves, transparency requirements for training data, and a collective licensing system with remuneration for authors. The text is not binding on anyone, but it does indicate the direction in which the Commission is being pushed.
Meanwhile, the first European trial is already underway. In case C-250/25, Like Company v Google Ireland, A Hungarian judge asked the Court of Justice whether the response of a chatbot that reproduces excerpts from articles constitutes communication to the public for the purposes of publishers' related rights, whether training constitutes reproduction, whether it falls within the exception of Article 4, and whether the provider is liable for what the model outputs at the user's request. These are the same steps the Department of Justice has attempted to keep separate.
Judge Sidney Stein will rule on fair use in a proceeding in which one of the parties now has the federal government on its side, with a decision that openly declares its intention to influence the outcome of a dispute between private individuals. This, rather than the fair use doctrine, will be worth returning to.
An earlier version of this article was published on September 3, 2026, on Fulvio Sarzana's blog on Il Sole 24 Ore.
English 
Italian 