Munich court found six songs sit inside Suno's model, then applied US law and rejected fair use

On 31 July 2026 the 42nd Civil Chamber of the Regional Court of Munich I, the chamber that specialises in copyright, largely upheld GEMA's claims against Suno for injunctive relief, disclosure and damages, in case 42 O 763/25. The court's own press release, issued the same day, sets out reasoning that is more specific than how the ruling has been summarised since.

The case concerned six well known works: Atemlos durch die Nacht by Kristina Bach, Rasputin by Frank Farian, Fred Jay and George Reyam, Big in Japan and Forever Young by Marian Gold, Bernhard Lloyd and Frank Mertens, the chorus of Mambo No. 5 by David Lubega and Christian Pletschacher, and Daddy Cool by Frank Farian. Infringement of the lyrics was not in issue. To generate the disputed outputs, GEMA's side entered the original lyrics of each song, the desired musical style and the title. The prompts specified nothing about melody, harmony, rhythm or arrangement.

The first finding is the one that has travelled. The court held that the works are reproducibly contained in versions v3.5 and v4 of Suno's models. It reached that by comparison: the works in the training data matched the renderings in the outputs closely enough that, given the length and complexity of the music, chance was excluded as an explanation. The court described this as memorisation, the situation where a model does not merely extract information from the training set but retains content of the training data in its post training parameters. Suno had argued the opposite, that the works were not protected in the first place, that they were not recognisable in the outputs, and that its weights and parameters represent mathematically learnt patterns and generalised features such as syntactic, semantic and contextual relationships, so that any similarity came from prompt induced narrowing of the search space and statistically learnt patterns. The court rejected that. Memorisation was held to infringe the reproduction right under section 16 of the German Copyright Act, and not to be covered by the text and data mining exception in section 44b.

The second finding concerns outputs. The court held the original elements of the works were recognisable in what Suno produced, and that Suno, not its users, was responsible, because the prompts were simple and open ended while Suno selected the training data, built the architecture and was responsible for the memorisation. On the court's reasoning the models determined the content of the outputs. Merely offering the model and the application for generating music was itself held to breach the unnamed right of public communication under section 15(2).

The jurisdictional holding is narrower and more technical than a general claim that a German court can reach any AI service whose outputs land in Germany. International jurisdiction over the acts committed on United States territory rested on section 131(1) and (2) of the German Collecting Societies Act, a provision the court said governs both local and international jurisdiction and creates a special venue of related subject matter for privileged collecting societies whose purpose is administering copyright. The German law findings rested on a different fact: the models were stored on servers in Germany.

The third point is the one that should unsettle American AI developers. Applying the lex loci protectionis principle, the court held that United States law governed the infringements committed in the United States, and then applied it. It found the copies made for training were not covered by fair use under 17 U.S.C. section 107, because the works reappeared in the outputs, and that every factor to be examined under the Supreme Court's Warhol framework told against Suno. The court expressly distinguished the American Bartz and Kadrey decisions, in which two United States courts treated the training of AI models as covered by fair use, on the ground that in those cases the training data was not, or not substantially, made accessible to users in the outputs. A German court has now applied American fair use doctrine and reached the opposite result on different facts.

How the training data was obtained also mattered. The court records that Suno used stream ripping techniques to extract and copy the works from YouTube, circumventing the rolling cipher, a technical protection measure the platform implemented to prevent the downloading of audio and video.

GEMA's chief executive, Tobias Holzmueller, said the chamber had made clear that AI models built on the theft of intellectual property are not protected by the legal order, and that providers must acquire licences. Its general counsel, Kai Welp, said it was astonishing that AI systems apparently store nearly complete works on a considerable scale.

Much remains open. The court's press release states the judgment is not legally binding. JUVE Patent reported that Suno intends to explore all options including an appeal. Damages were established in principle but not quantified, which is what the disclosure order is for. And a reference pending before the Court of Justice of the European Union in the Like Company case, on which Licentium reports a ruling was expected in September 2026, could still change how Article 4 of the copyright directive, the European source of the text and data mining exception, is read.