AI training copyright: US backs fair use, Munich rules against Suno

Within five weeks of each other, a German court and the US government took opposite positions on a question that matters to every organisation using generative AI: can a model be trained on copyrighted work without a licence? On 31 July 2026 the Munich Regional Court I ruled against the US music generator Suno in a lawsuit brought by GEMA, the German music collecting society. On 1 September the US Justice Department told a federal court in New York that training large language models on copyrighted text is fair use.

Neither development settles the law. The Munich judgment is a first-instance decision that is not final, and Suno has said it is considering its options, including an appeal. The Justice Department filing is a statement of interest in the litigation against OpenAI. It sets out the government's view, but it is not a ruling, and the question it addresses remains for the court.

Read together, the two show where legal risk is moving. The US government argues that training should be judged separately from what a model later produces. The Munich court looked at what Suno's models could reproduce and treated that as evidence that the works were inside them. For teams deploying generative tools, exposure increasingly depends on memorisation, on outputs, and on where those outputs reach users.

How copyright law meets model training

Training a model involves copying at several stages, from collecting material and preparing datasets to running that data through training. In both the United States and Germany, reproduction is a right reserved to the rights holder, so disputes turn on which exceptions or defences cover which copies.

The United States relies on fair use, a defence that courts assess case by case through four factors: the purpose and character of the use, including whether it is transformative; the nature of the work; the amount taken; and the effect on the market for the original. The Justice Department's filing concentrates on the first and fourth factors and deals with the other two only briefly.

German courts work instead through exceptions written into statute. The one that matters most for AI is the text and data mining exception in Section 44b of the German Copyright Act, which implements Article 4 of the EU's Digital Single Market Directive. Rights holders can opt out of it in machine-readable form, and a Hamburg court in Kneschke v LAION took a rights-holder-friendly view of what counts as a valid opt-out.

The concept linking both systems is memorisation. The Munich court described it as a model encoding specific training content in its parameters so that the content can later be reproduced in whole or in part, as distinct from learning patterns or correlations. If a work is memorised in that sense, the model itself can be treated as containing a copy, whatever numerical form it takes.

What the Munich court decided

GEMA sued on behalf of the composers of six works: Atemlos durch die Nacht, Rasputin, Daddy Cool, Big in Japan, Forever Young and the refrain of Mambo No. 5. The case, numbered 42 O 763/25, concerned the musical compositions rather than lyrics. According to the court's findings as reported by DLA Piper, the recordings used for training had been ripped from YouTube in a way that circumvented its technical protection measures.

On jurisdiction, the court relied on a provision of the German Collecting Societies Act that gives collecting societies a forum extending to related infringements abroad. On applicable law it followed the territoriality principle, judging training carried out in the United States under US law and acts in Germany under German law. Bristows describes the result as the first European ruling that training carried out entirely outside the EU can create liability in Germany.

Applying US law, the court treated the training copies as reproductions under the US Copyright Act and rejected fair use. It accepted that training can in principle be transformative, but found the circumstances weighed against Suno: outputs were substantially similar to the originals, the data had been obtained by circumvention, and the service could substitute for the works. It distinguished the US decisions in Bartz v Anthropic and Kadrey v Meta, where outputs had not been shown to closely reproduce training works and market harm had not been established.

Under German law, the court found the works memorised in Suno's models, reportedly versions 3.5 and 4, and held that this was a reproduction even though the works existed only as parameters. The text and data mining exception could cover copies made in preparation for analysis, but not reproductions that persist in the trained model. The court also rejected the argument that meeting obligations under Article 53 of the EU AI Act makes further licensing unnecessary, holding that the Act does not change copyright law.

Outputs were treated as a further infringement. Because the test prompts were simple and open-ended, giving only a title, lyrics and a style description, the court attributed recognisable outputs to Suno rather than to its users. It granted four injunctions, covering training on the works in the United States, storing them in the models, offering the trained model in Germany and generating infringing adaptations, and upheld claims for information on several of the works, which GEMA can use to prepare a damages claim.

What the US government argued

The Justice Department filed its statement of interest on 1 September 2026 in the consolidated copyright litigation against OpenAI before Judge Sidney Stein in the Southern District of New York. The filing addressed claims brought by The New York Times, and the department said its reasoning also applies to related cases brought by book authors and publishers. The Intercept, another plaintiff in the consolidated proceedings, reported on the filing on 2 September.

On the first factor, the government argued that training is highly transformative because a model does not use an article to inform or entertain readers in the way its author intended. On the fourth, it said training on its own makes no protected expression available to the public, so it does not replace an article in the legally relevant sense. It favoured the reasoning in Bartz over Kadrey, rejected the market dilution theory applied in Kadrey, and said the reasoning of the Register of Copyrights did not warrant deference.

The policy case rested on competition and security. The department argued that requiring licences for training could create entry barriers that only the largest technology companies could absorb, and described domestic AI development as important to economic growth and national security. The Intercept placed the filing within the administration's wider effort to protect US AI companies in a race with rivals abroad, most notably in China. Matt Topic, the lawyer representing The Intercept, called the position out of touch and a threat to the financial survival of news organisations.

The limits of the filing matter as much as its argument. The department said it was addressing only the use of works during training, and it left open questions about how training data is acquired and stored, and about outputs that reproduce protected expression. According to Sterne Kessler, it also sidesteps whether training on pirated material can be fair use. The government accepted that outputs which reconstruct copyrighted material could raise different issues.

A fair use argument for training offers little protection to a model that can be prompted into reproducing the works it learned from.

What each development does and does not decide

The Munich judgment is a first-instance decision in one dispute over six works. It is not final, and DLA Piper notes that the Higher Regional Court of Munich or the Federal Court of Justice may have to rule on the issues later. The chamber saw no need to refer questions to the Court of Justice of the EU, although a referral remains possible, and no damages figure has yet been set.

Its treatment of US fair use is a German court applying foreign law to particular facts: data obtained by circumventing technical protections, and outputs close to the originals. It is not a precedent that US courts must follow. It also says little about training on lawfully accessed material that a model does not memorise, since the court accepted that preparatory copies made for analysis may fall within the text and data mining exception.

The Justice Department filing is advocacy, not law. It asks the court to adopt a view of training that the court may or may not accept, and by its own terms it does not cover acquisition, storage or infringing outputs. The claims in the consolidated case also differ: The Intercept's surviving claim against OpenAI is under the Digital Millennium Copyright Act, after the judge then presiding dismissed its claims against Microsoft in November 2024.

What remains disputed

The central dispute is how to treat memorisation. The Munich court sees recognisable reproduction as proof that a work sits inside the model. The US government treats a small number of reconstructive outputs as anomalies that should not justify broad restrictions on training. Both views can hold at once in different legal systems, which is why they matter for any service offered in more than one market.

Several open cases will shape the answer. GEMA's earlier Munich judgment against OpenAI, handed down on 11 November 2025 over song lyrics, is under appeal. The Court of Justice of the EU is due to consider memorisation in a case brought against Google by the Hungarian publisher Like Company. In the United States, the consolidated OpenAI litigation continues, and Sterne Kessler describes the Bartz and Kadrey decisions as conflicting.

Licensing is the unresolved commercial question underneath. DLA Piper suggests that operators of generative models in the EU consider acquiring rights to training data where it could reappear in recognisable form in outputs. The US government argues the opposite policy case: that requiring licences for training would favour the largest companies. Because the German judgments are not final and the US filing is not a ruling, all of these positions should be treated as developing.

What this means for organisations deploying generative AI

Many organisations that deploy generative AI never train a model themselves, but they do publish what models produce. The practical risk therefore sits in vendor choice, output controls and contracts.

• Map where generated content reaches users: a model defended under a US fair use argument can still face injunctions in Europe if it memorises and reproduces protected works there.
• Ask vendors how training data was obtained, including whether any was collected by circumventing technical protections, and how they handle machine-readable rights reservations.
• Ask for evidence of memorisation testing, such as whether simple prompts using a title, lyrics or a style reproduce recognisable works, and whether that testing is repeated after model updates.
• Review indemnities and warranties, noting that the Munich court attributed outputs from simple prompts to the provider rather than the user; that allocation may differ for more elaborate prompting.
• Add output filters and human review for content that could reproduce third-party works, particularly music, news text and other media published under the organisation's own name.
• Do not treat compliance with the EU AI Act's obligations for general-purpose models as a copyright defence; the Munich court held that the Act leaves copyright law intact.