Munich court: a model that memorises songs is copying them

On 31 July 2026 the Munich Regional Court largely ruled for GEMA, Germany's music collecting society, against Suno. The 42nd civil chamber, which specialises in copyright, granted most of GEMA's claims for an injunction, disclosure and damages over six songs, according to the court's press release and the 137-page judgment (case 42 O 763/25). The judgment is not final and can be appealed.

Whose songs were in the case?

Six works from GEMA's repertoire: "Atemlos durch die Nacht" (Kristina Bach), "Rasputin" (Frank Farian, Fred Jay, George Reyam), "Daddy Cool" (Frank Farian), "Big in Japan" and "Forever Young" (Marian Gold, Bernhard Lloyd, Frank Mertens), and the refrain of "Mambo No. 5" (David Lubega, Christian Pletschacher). The case was about the music only. The lyrics were not part of it.

How GEMA got the right to sue matters to writers. The judgment records that between September 2024 and October 2025 the writers (for Farian and Jay, family members or a former manager) and their publishers signed add-on agreements. These gave GEMA the exclusive right to license their works for AI training and AI output. Frank Mertens, the third writer of "Big in Japan" and "Forever Young", didn't sign one. For those two songs GEMA asked only for an injunction, on an authorisation from Gold and Lloyd. Disclosure and damages cover the other four.

What did the court decide?

Memorisation is copying. According to the facts set out in the judgment, Suno trained in the US on a corpus that included millions of complete recordings. It got these songs by stream-ripping them from YouTube, getting around a protection YouTube uses to stop downloads. The court found the six works were reproducibly contained in Suno's v3.5 and v4 models, which were stored on servers in Germany. It called this memorisation: the model doesn't just learn from its training data but takes content from it into its parameters. Given how long and complex the songs are, the court ruled out chance. That memorisation is a reproduction under §16 of the German Copyright Act, and the text-and-data-mining exception doesn't cover it.

Suno, not the user, made the outputs. The court found recognisable original elements of the songs in Suno's outputs. GEMA's prompts gave the original lyrics, a genre and the title, with nothing about melody, harmony, rhythm or arrangement. Because Suno chose the training data, built the models and caused the memorisation, the court held the models, not the prompts, decided what came out. Offering the model and the app was itself a communication to the public.

US fair use, decided in Munich. The court took jurisdiction over the US training using a special rule for collecting societies (§131 of the German Collecting Societies Act). It then applied US law to the copying done in the US. It distinguished Bartz v. Anthropic and Kadrey v. Meta, where US courts found training to be fair use but the works didn't substantially reappear in outputs. The judgment even accepts that training a music generator can, in general, be transformative. Here it wasn't, because the outputs were substantially similar to the songs. The injunction covers the US training copies as well as the model in Germany. Independent summaries from WIPO Lex and Center Consulting agree on the outline.

What the headline leaves out

It isn't the first memorisation ruling. In November 2025 the same chamber held that OpenAI's GPT-4 and 4o models had memorised the lyrics of nine German songs, and that this counted as reproduction. OpenAI has appealed to the Higher Regional Court in Munich, ZDF reports. What is new now is audio: melody, harmony and rhythm in a music generator, plus a German court ruling on US fair use.

The proof took work. The judgment counts 3,779 prompts by GEMA's staff and lawyers. 338 of them concerned these songs, including 176 identical prompts for "Atemlos" and 124 for "Big in Japan". The court said repeating an identical, open prompt doesn't steer the model. It still turned down GEMA's claim under the making-available right, because a song that can take up to 176 tries to appear isn't available to users at a time of their choosing. GEMA won on the general right of communication to the public instead. The court also noted that GEMA didn't file the other infringing outputs it said it had for these songs.

Suno told ZDF it disagreed with the ruling, which it said rested on a wrong picture of the technology, and that its models are trained to create new songs, not reproduce old ones. It is weighing all options, including an appeal. In GEMA's statement, CEO Tobias Holzmüller said AI built on stolen intellectual property isn't protected by law. GEMA says it sued in January 2025 after Suno didn't respond to a licensing request.

What we think

Memorisation is the technical fact that breaks the "it only learns patterns" defence. That defence describes what a model does on average. This court ruled on what one model did with six specific songs, and it put the risk on the company that built the model. Even under US fair use, the line it drew is the one artists care about: training can be fine in general, but not when the songs come back out.

The less obvious lesson is about paperwork. GEMA could enforce because its writers had signed AI-rights add-ons, and where one co-writer hadn't, the damages claim for two songs fell away. Individual songwriters can't run 338 prompts and a 137-page case; a collecting society can. So if you're a songwriter, find out whether your society, or your publisher, actually holds AI rights in your works. If you hand them over, ask how licence money from AI firms will be split and reported back to you. Enforcement only helps writers if the money reaches them.

Expect two industry responses. First, more work on stopping models from regurgitating songs: output filters and training methods that reduce memorisation. GEMA argued in court that Suno's existing safeguards only target regurgitation, not memorisation. Second, a price on repertoire held by societies that can sue in Europe. Because the court reached US training through GEMA's standing, moving servers out of Germany doesn't obviously remove the risk. Leaving the repertoire out does. If that happens, the cost of doing this properly could be models that know less European music. The right answer is licences that pay writers, not quiet exclusion.

This is still a first-instance ruling. The OpenAI appeal comes first, and ZDF notes that a pending EU Court of Justice ruling in a case involving Google could arrive by the end of the year. Treat Munich as a strong signal, not settled law.

Sources

• Landgericht München I, Pressemitteilung 16 (2026): GEMA v. SUNO
• Landgericht München I, Endurteil 42 O 763/25, 31 July 2026 (PDF)
• Landgericht München I, Pressemitteilung 11 (2025): GEMA v. OpenAI
• WIPO Lex: GEMA v Suno
• Center Consulting: GEMA v. Suno Munich ruling
• miz: GEMA statement on the Suno ruling
• ZDFheute: Gema gewinnt gegen Suno