Joris DeeneIn the United States, it is—at least according to the U.S. Department of Justice: According to the U.S. government, training a language model on copyrighted texts constitutes “fair use” under Section 107 of the Copyright Act, because the copying serves to learn language patterns and the work itself is not disclosed to the public. In Belgium and the European Union, the situation is different: there, training constitutes a reproduction that is permitted only under the exception for text and data mining, unless the copyright holder has objected in a machine-readable format. The reason for this is a statement of interest which the Department filed on September 1, 2026, with the federal court in New York, in the consolidated case brought by The New York Times and other publishers and authors against OpenAI (U.S. District Court, Southern District of New York, 25-md-3143, document 316).
In short:
- On September 1, 2026, the U.S. government explicitly asks the court to rule that training language models on protected texts does not constitute infringement, and rejects the theory of “market dilution” from the Kadrey v. Meta decision of June 25, 2025, as fundamentally flawed.
- The article is not a ruling: the judge is not bound by it; The New York Times called it a choice in favor of “a handful of trillion-dollar AI companies,” and Judge Stein asked the newspaper to explain by September 11, 2026, why its case should not be stayed.
- For a Belgian company, the legal framework remains unchanged: Article XI.190, 20° of the Code of Economic Law permits training as text and data mining, but a machine-readable opt-out by the rights holder overrides that exception, and the AI Regulation requires providers to respect that opt-out.
Why the U.S. government is intervening in a dispute between private parties
On December 27, 2023, The New York Times filed a complaint against Microsoft and OpenAI (U.S. District Court, Southern District of New York, 1:23-cv-11195). The newspaper accuses OpenAI of copying millions of articles to train the GPT models, and claims that ChatGPT can reproduce passages from those articles. Since April 2025, that case and a dozen similar lawsuits filed by publishers, book authors, and, among others, writer Ta-Nehisi Coates have been consolidated before Judge Sidney Stein.
A statement of interest is a procedural document through which the U.S. Department of Justice explains the interests of the United States in an ongoing case without becoming a party to the case itself. The legal basis is 28 U.S.C. § 517, which imposes no deadline and requires no court approval. It is similar to an amicus curiae brief, but with one difference: in this case, it is not an interest group or an academic speaking, but the federal executive branch itself. The Department emphasizes that it is not suggesting that OpenAI acted with the government’s consent or for the government’s benefit.
The interests cited by the ministry are remarkably broad. It refers to the presidential decree of January 23, 2025, regarding the Removing Barriers to U.S. AI Leadership, to a Decree of June 2, 2026 on AI innovation and safety, and toward a White House Policy Framework of March 20, 2026 which states that training AI models on protected material does not, in and of itself, constitute a copyright infringement. The common thread is national security: according to the department, rules that hinder the development of a U.S. AI industry give an advantage to foreign adversaries who are not hindered by such rules.
There is also a competition-related argument. If AI training were permitted only under a license, the ministry argues that only the largest technology companies would be able to afford the licensing fees. The result would be an oligopoly in AI training, and the licensing fees would primarily benefit the established media companies with the largest archives. The ministry refers to “large subsidies for established mainstream media companies.” In a footnote, it adds that it does not take a position on whether a licensing system would be financially or logistically feasible, and that, regardless of this, publishers already enter into licensing agreements for access to current and paid content.
What the U.S. government is arguing before the court
Fair use is the U.S. copyright exception: Section 107 of the Copyright Act Allows the court to determine on a case-by-case basis whether a use is fair, based on four factors: the purpose and nature of the use, the nature of the protected work, the amount of the reproduced portion, and the effect on the market for the work. The ministry bases its argument on the first and fourth factors.
According to the ministry, the training is “spectacularly” transformative
With regard to the first factor, the ministry concurs with the decision in the case of Bartz v. Anthropic (U.S. District Court, Northern District of California, June 23, 2025, 3:24-cv-05417), in which Judge Alsup called the training of language models on books “transformative—spectacularly so.” The model does not use the text to repeat the expression, but to learn statistical patterns in language: vocabulary, sentence structure, and knowledge. The purpose of the copy (building an interactive model) differs in nature from the purpose of the copied work (informing or entertaining a readership).
To that end, the ministry relies on the Google v. Oracle ruling (Supreme Court (U.S.) April 5, 2021, 18-956), in which the Supreme Court accepted the literal copying of 11,500 lines of programming code as fair use because it allowed programmers to apply their existing knowledge in a new environment. If that is transformative, the department argues, then so is the copying of text to train a model to recognize relationships between data.
A key point in the argument is the distinction between training and output. Since the Warhol v. Goldsmith decision (Supreme Court (U.S.) May 18, 2023, 21-869) the fair use test must be applied to each individual use. According to the ministry, the fact that a model may, in exceptional cases, reproduce portions of an article is a matter concerning the output, which must be assessed separately and does not detract from the transformative nature of the training. Even if some outputs were to infringe copyright, any sanction must be limited to what is necessary to fully compensate the plaintiffs; a “small proportion of anomalous reconstructive outputs” does not justify a ban on training as such.
Market distortion requires substitution, and according to the ministry, that is not available
For the fourth factor, the ministry cites the Authors Guild v. Google ruling (U.S. Court of Appeals for the Second Circuit, October 16, 2015, 804 F.3d 202), the ruling that recognized the scanning of millions of books for a search function as fair use. Market harm is only taken into account if the use serves as a substitute for the original and results in “significant substitute competition.” According to the ministry, training does not reveal anything to the public: it makes a copy to train a model, and nothing more. Even outputs that compete with newspaper articles without substantially resembling them are not substitutes in the copyright sense, because a genre or subject matter is not protected.
The most scathing part of the article is directed against the ruling in the Kadrey v. Meta case (U.S. District Court, Northern District of California, June 25, 2025, 3:23-cv-03417). Judge Chhabria did rule in Meta’s favor on that point, but only because the plaintiffs had failed to present evidence of market harm. In his reasoning, he accepted a theory of “market dilution”: a language model trained on books could flood the market with competing books and thus undermine the incentive to write, even if no single output resembles a single original work. The Ministry calls this reasoning an obiter dictum—made without debate—and “deeply flawed”: it conflates training and output into a single use and overlooks the most fundamental element of an infringement, namely the substantial similarity between the works.
The ministry illustrates this with a literary anecdote. As a teenager, Joan Didion retyped Hemingway’s stories to learn how his sentences worked. According to Kadrey’s logic, the article states, she would have owed Hemingway a fee for every publication, because her learning process and her later work would constitute a single use that competes with other authors. The report from the U.S. Copyright Office dated May 2025 According to the ministry, the report on AI training—which embraced a similar dilution theory—carries no weight whatsoever. The document dryly notes that the director of that agency is currently challenging her dismissal in court.
Legal analysis and interpretation
The document is a policy paper disguised as a legal brief, and the judge knows that
Anyone reading the statement will notice that the legal arguments don’t begin until page six. The five pages preceding it deal with national security, weapons systems, competition with foreign countries, and the interests of independent media. This is no coincidence: the ministry is presenting the court with a policy choice and framing it within existing case law. The New York Times’ response was therefore immediate and political: according to the newspaper, the government is choosing a group of a handful of AI companies at the expense of the creators whose work has been stolen.
Legally speaking, the document does little to alter the balance of power. A statement of interest is not binding on the judge, and Judge Stein has shown over the past two years that he sets his own pace. Moreover, the arguments are not new: they stem from the Bartz and Kadrey decisions and the Authors Guild ruling, and OpenAI has already raised them itself. What is new, however, is that the federal government is weighing in for the first time on the central question in the dozens of AI copyright cases pending in the United States. That weight is real, but it is the weight of a party’s position, not that of a precedent.
There is also a dissenting opinion within the U.S. judiciary itself that the Department fails to mention. In February 2025, a judge in Delaware rejected the fair use defense of an AI developer who had used legal summaries from a competitor, as we discussed in our blog post on the question of whether there is not fair use for training an AI model. That case involved a non-generative system and a direct competitor, but it shows that the fourth factor can indeed work against an AI developer once the model replaces an existing market. The Third Circuit Court of Appeals heard the case on June 11, 2026 and still has to decide.
The distinction between training and output is less clear-cut in Europe
The entire structure of the ministry rests on a single premise: training does not reveal anything, so training cannot cause market harm. In Europe, that premise is under pressure, and this is due to the way the right of reproduction is interpreted here. The Munich I Regional Court ruled in the case of the German collecting society GEMA v. OpenAI (LG Munich I, November 11, 2025, 42 O 14139/24) that the memorization of song lyrics within the parameters of a language model constitutes a reproduction in and of itself, even if it cannot be found as a file. According to the court, the exception for text and data mining did cover the creation of the dataset, but not the storage of the works within the model. On July 31, 2026, the same court prohibited the AI music generator Suno from memorizing and reproducing well-known hits, as we analyzed in our blog post on the question of whether a AI music generators may be trained on existing hits.
The contrast is stark. While the ministry states that training “does not reveal anything to the public,” the German court holds that a model capable of reproducing a work embodies that work. The American reasoning focuses on purpose and the market; the European reasoning focuses on the technical act itself. A rights holder in Belgium who can demonstrate that a design recognizably reproduces his work therefore does not need to engage in a debate about market dilution: he simply shows a reproduction and compels the developer to invoke an exception.
In Belgium, the question is not “Is this fair use?” but “Is the opt-out valid?”
Belgian and European copyright law does not recognize an open “fair use” exception. The exceptions are set forth in an exhaustive list, and training falls under only one of them. According to Article I.13, 10° of the Code of Economic Law, text and data mining is an automated analysis technique used to discover patterns, trends, and correlations in digital text and data. Article XI.190, 20° of the Code of Economic Law (CEL) allows the reproduction of lawfully accessible works for text and data mining, provided that the author has not expressly reserved such use in an appropriate manner. For content available online, that reservation is considered appropriate only when machine-readable means are used. That provision implements Article 4 of the DSM Directive as of April 17, 2019. For research organizations and cultural heritage institutions, there is also a broader exception without an opt-out in Article XI.191/1, § 1, 7° of the WER.
The European legislature has confirmed that choice in the AI Regulation. Article 53, paragraph 1, (c) of the AI-relateddmeeting As of June 13, 2024, providers of general-purpose AI models are required to implement a policy that complies with Union copyright law, in particular to identify and respect the copyright notices set forth in Article 4 of the Directive, and point (d) requires them to publish a summary of the training content used. Recital 106 adds that this obligation applies to any provider placing its model on the EU market, regardless of where the training took place. An American model trained in the United States under fair use must therefore still be able to demonstrate in Europe that it has respected the European opt-outs.
It is this key issue that is shifting the debate in Europe. On December 10, 2025, the Hanseatic Higher Regional Court upheld a Judgment of the Hamburg Regional Court on September 27, 2024 (310 O 227/23), in which it accepted the creation of a training dataset by the LAION research association as text and data mining, and ruled that appeal to the Supreme Court We discussed that issue in our blog post about whether AI Can Use Your Online Photos for Training Without Permission: A prohibition stated in plain text in the terms and conditions is not sufficient; the opt-out must be technically readable. You can read more about the limits of this exception in our blog post on the question When AI training violates the three-step test, and the broader context can be found on our page about Artificial Intelligence and Copyright.
The decisive European ruling has not yet been issued. The Court of Justice is hearing the case C-250/25, Like Company v. Google, raised for the first time the question of whether training a chatbot constitutes reproduction, whether the exception for text and data mining covers such training, and whether outputs that reproduce excerpts from press publications constitute a communication to the public. The Grand Chamber held a six-hour hearing on March 10, 2026; the Advocate General’s opinion was scheduled for September 3, 2026, and the judgment will follow thereafter. The European Copyright Society warned the Court in a Opinion from March 2026 for a hasty response, because the preliminary questions confuse chatbots, language models, and search engines. The outcome of that case carries more weight for a Belgian company than what the U.S. government is arguing in New York.
The competition argument cuts both ways
The ministry warns of an oligopoly: only the wealthiest companies can afford the licenses. That argument warrants a caveat. In 2025, Anthropic paid $1.5 billion to settle a class-action lawsuit filed by authors—about $3,000 per book for approximately 500,000 works; the settlement was approved on July 20, 2026: Final Approval. That case did not concern the training itself—which Judge Alsup had ruled constituted fair use—but rather the downloading of seven million books from pirate sites. It illustrates two things. First, the origin of the training data remains a separate risk, even in the United States: fair use covers the learning, not the illegal source files. Second, the amounts already involved in licenses and settlements are of a magnitude that, in fact, only the largest players can bear. The ministry therefore has a valid point regarding the barrier to entry, but this point applies equally to the rights holders: without a licensing requirement, the cost of the system shifts entirely to the creators, and the ministry itself acknowledges that it does not take a position on the feasibility of a licensing market.
Specifically, what does this mean?
For publishers, authors, and media companies in Belgium. The U.S. position does not affect your rights, but it serves as a reminder that a U.S. judge may treat your content differently than a Belgian judge would. Your protection in Europe depends on a single action: a valid, machine-readable copyright notice on everything you post online. A single sentence in your terms and conditions may not be sufficient. If you want to verify whether your current opt-out is sufficient against the major model developers, the technical and legal validity of that notice is the first thing we’ll check for you. In addition, it’s worth documenting the outputs: a model that recognizably reproduces your articles constitutes a reproduction in Europe, a fact that German case law has now recognized twice.
For Belgian developers of AI models and applications. Don't count on fair use, even when you further train or refine U.S. models using your own data. The questions you must be able to answer are European in scope: Did you have lawful access to the works? Did you respect the opt-outs? How long do you retain the copies? And do you have the summary of training content required by Article 53 of the AI Regulation? The research exception under Article XI.191/1, § 1, 7° of the WER is broader but is available only to research organizations and cultural heritage institutions, and the German LAION case law regarding what constitutes “research” is still being reviewed by the Bundesgerichtshof. Anyone licensing protected content would be well advised to ensure that the license explicitly states that it covers training and professional development, and, in case of doubt, to seek assistance from a attorney specializing in copyright law.
For companies that use generative AI. Under European case law to date, liability for infringing output lies with the provider of the model, not with the user who simply entered a prompt. That does not relieve you of responsibility: anyone who publishes generated text that is clearly derived from an existing article is themselves committing an act of reproduction and communication to the public. An internal policy on the use of AI-generated content and monitoring for verbatim copying remain necessary. The copyright exceptions that you can invoke in Belgium, such as the right to quote, are described on our page about the exceptions to copyright.
Frequently asked questions (FAQ)
Can an AI model be trained on copyrighted texts?
In the United States, according to the Department of Justice and two federal judges, this constitutes fair use as of 2025, but the matter is still pending before a judge in New York and on appeal elsewhere. In Belgium, training constitutes a reproduction that is permitted only as text and data mining under Article XI.190, 20° of the Belgian Copyright Act (WER): the works must be lawfully accessible, and the copyright holder must not have objected in a machine-readable format. If the copyright holder has objected, a license is required.
What is a statement of interest, and is the judge bound by it?
A statement of interest is a document in which the U.S. Department of Justice, pursuant to 28 U.S.C. § 517, explains the interests of the United States in a pending case, without becoming a party to the case and without the court’s permission. It is comparable to an amicus curiae brief. The judge is not bound by it: he considers it as a reasoned opinion, alongside the arguments of the parties. In the case against OpenAI, Judge Sidney Stein will decide independently on the fair use issue.
How can I protect my website or publications from being used for AI training?
Through a machine-readable copyright notice, as required by Article XI.190, 20° of the WER for online content. In practice, this involves technical signals that AI crawlers can read, combined with an explicit notice in your site’s legal disclosures. A prohibition in plain text alone is not sufficient, as German courts ruled in the LAION case. Providers of general-purpose models are required under Article 53 of the AI Regulation to respect that reservation, even when they train their models outside of Europe.
Does fair use apply in Belgium as well?
No. Belgian and European copyright law operates under an exhaustive list of exceptions, of which text and data mining is one. A judge cannot create a new exception simply because a particular use appears to be socially beneficial, unlike a U.S. judge, who is permitted to do so under Section 107 of the Copyright Act. Arguments regarding national security or innovation, such as those put forward by the U.S. government, therefore play no role in Belgian infringement proceedings; only the conditions of the statutory exception matter.
Conclusion
The Statement of Interest dated September 1, 2026, marks the first time the U.S. federal government has weighed in on the copyright status of AI training, and it does so without reservation: training constitutes fair use, and the “market dilution” theory from the Kadrey decision is flawed. The judge in New York is not bound by this, and the document does not alter Belgian law. Here, training constitutes a reproduction that falls under the text and data mining exception of Article XI.190, 20° WER, with the machine-readable opt-out as the cornerstone and the AI Regulation as the enforcement tool, and with the Court of Justice’s ruling in the Like Company case as the real turning point that is yet to come.
ICT Rechtswijzer is the knowledge platform of the intellectual property, IT law, AI law, data protection, and media law department at Everest Advocaten, led by Joris Deene, Esq. Our attorneys assist publishers and authors in establishing a valid reservation of rights against AI training and in documenting infringing outputs, and assist AI developers in complying with the exception for text and data mining and the obligations under the AI Regulation.
Dutch 
English