A German Court Applied US Fair Use to Suno. It Still Lost.

TL;DR

On July 31 the 42nd Civil Chamber of the Regional Court of Munich I ruled that Suno infringed copyright in six songs from the repertoire of GEMA, Germany's state-mandated music collecting society. Case number 42 O 763/25. Two findings travel a long way past music. First, the court held the songs were retained inside the model weights in reproducible form, which makes them copies rather than text and data mining. Second, and this is the part American labs should read twice: because the training copying happened on US soil, the court applied US copyright law to it, found no fair use, and expressly distinguished the Bartz and Kadrey decisions on the ground that Suno's outputs handed the songs back. Suno must stop, disclose the revenue it earned, and pay damages to be set in a later proceeding. The judgment is not yet enforceable, and Suno is weighing an appeal.

What the court actually held

Presiding judge Elke Schwager sided with GEMA on nearly every point. The six works at issue were "Daddy Cool" and "Rasputin" (Frank Farian and co-writers), "Forever Young" and "Big in Japan" (Alphaville), "Atemlos" (Kristina Bach), and "Mambo No. 5" (David Lubega, Christian Pletschacher). Lyrics were not part of this case. This was about melody, harmony and rhythm, which is a harder thing to prove and a much harder thing to explain away.

The chamber found two separate infringements stacked on top of each other. The reproduction right was infringed because, per JUVE Patent, "the disputed songs were effectively stored within the defendant's AI music models, which were hosted on servers in Germany." The right of public performance was infringed separately, by offering the model and its outputs to the public. Music Ally reports the court found the works reproducibly contained in Suno's v3.5 and v4 models.

Suno's attempt to push liability onto its own users went nowhere. The court noted that the users "only entered basic prompts, such as lyrics and a musical style. The defendant designed, trained and operated the models."

The weights are the copy

Suno's core defence was the one every lab makes: the model learns patterns and generates new songs, it does not keep old ones. The chamber rejected it on the evidence. At the March hearing GEMA played originals and Suno outputs back to back, and the court concluded that "given the complexity and length of the musical pieces, randomness as the cause of their playback can be ruled out."

That is the whole doctrinal move, compressed: memorization is reproduction. The text and data mining exception, the court held, "does not apply, because the works were not only analysed during training but retained in the models in a reproducible form."

If you want the intuition, a memorizing model is less like a student who studied the songbook and more like a very lossy archive with a natural-language search box bolted on. You cannot browse the file listing, nothing is stored as an MP3, and yet the right prompt still pulls a track out close enough that a judge recognises it. German copyright law does not care about the file format. It cares whether the work came back.

This is the same chamber that ruled against OpenAI on November 11, 2025 (case 42 O 14139/24), finding that ChatGPT had memorized nine sets of German song lyrics and that encoding them in model parameters was a reproduction under Section 16 of the German Copyright Act. Note that "Atemlos" appears in both cases. Kristina Bach is now two for two against frontier AI companies in the same Munich courtroom, which is a better win rate than most litigation practices.

Munich ran America's fair use test, and Suno failed it

Here is the finding with the longest reach. The copying during training happened in the United States, so the court applied US copyright law to that act, and then held that the copies and outputs were not protected by fair use.

It got there by distinguishing the two 2025 California decisions that AI companies have been citing all year. In Bartz v. Anthropic and Kadrey v. Meta, US courts found training to be fair use largely because the original training materials were not reproduced for users in the outputs. Suno's were. Same doctrine, opposite result, and the hinge was not how the data was acquired but what the model emitted.

Strip out the jurisdictions and you get a rule of thumb worth pinning above your eval dashboard: fair use has been surviving where the model does not hand the source back. Every memorization bug you ship is a step toward the other side of that line. The German-hosted servers are what gave Munich its hook on the storage and output side, so "we train in California" is not the shield it looked like last summer.

Where German courts have drawn the line

Read this ruling next to the other German decisions and a fairly coherent map appears, one that splits the pipeline into stages rather than blessing or banning "AI training" as a whole.

In Kneschke v. LAION, the Higher Regional Court of Hamburg upheld the dismissal of the photographer's claims on December 10, 2025 (case 5 U 104/24), holding that downloading and processing images to build a dataset fell inside the TDM exception. Per Morgan Lewis, the Hamburg court also held that an effective opt-out has to be machine readable, not a sentence of legalese in your terms of service. That case is now headed to Germany's Federal Court of Justice, docket I ZR 281/25, with a hearing set for September 3, 2026. LAION won on dataset construction; Suno lost on what ended up in the weights.

What this changes for you

• Memorization is now a legal test surface, not a research curiosity. If a rightsholder can prompt your model into something a judge recognises, that transcript is the evidence. Extraction testing belongs in your release gate next to your safety evals.
• Machine-readable opt-outs are load-bearing. Article 53(1)(c) of the EU AI Act already requires general-purpose model providers to run a copyright policy that identifies and complies with reservations under Article 4(3) of the DSM Directive, "including through state-of-the-art technologies." Munich just showed what the downside looks like when you do not.
• Where you serve is part of your exposure. The German-hosted models are what anchored the storage and output findings. Inference geography is now a legal parameter, not just a latency one.
• Output-side controls are the cheapest insurance available. Dedup, near-duplicate detection at generation time, and an actual memorization eval cost far less than a disclosure order over your revenue.

The caveats, stated straight

This is a first-instance judgment. JUVE Patent reports it is not yet enforceable, and Suno has said it will explore all options including an appeal, telling Music Ally that the ruling "rests on a fundamental mischaracterization" of its technology. Nothing here is settled German law until the Higher Regional Court, and probably the Federal Court of Justice, has weighed in. GEMA's parallel win against OpenAI is still on appeal.

The cross-border move is also the most obviously appealable part of the decision. A German regional court running the US fair use four-factor analysis and reaching a different answer than two California federal judges is exactly the sort of thing an appellate bench likes to revisit.

And the scope is narrow on paper: six works, not GEMA's whole repertoire, with damages still unquantified pending Suno's revenue disclosure. Music also plays to the plaintiff here, because a melodic hook is short, distinctive and easy to A/B against an output in a courtroom. A text model's memorization exposure is real but harder to demonstrate to a judge in an afternoon.

Key Takeaways

• The Regional Court of Munich I ruled on July 31, 2026 (case 42 O 763/25) that Suno infringed six GEMA-represented songs, finding two separate violations: reproduction in the weights and making the works available through outputs.
• The TDM exception was held not to cover works "retained in the models in a reproducible form." Analysis is exempt; memorization is not.
• The court applied US copyright law to the US-side training and found no fair use, distinguishing Bartz v. Anthropic and Kadrey v. Meta because in those cases the outputs did not reproduce the source material.
• Germany's emerging line: dataset construction is TDM-covered (LAION, Hamburg), retention in the weights is not (GEMA, Munich), and the output is where you get caught.
• Remedies are an injunction, revenue disclosure, and damages to be determined. The judgment is not yet enforceable and an appeal is likely.
• The Federal Court of Justice hears the LAION appeal on September 3, 2026 (I ZR 281/25), which is the next real checkpoint for anyone training on European rightsholders' work.

Sources: GEMA press release, 31 July 2026, JUVE Patent, Music Ally, Landgericht München I press office (case 42 O 763/25), The Next Web, Morgan Lewis, AI and Copyright: Judicial Landscape in Germany, CMS on GEMA v. OpenAI, EU AI Act Article 53