On 31 July 2026, the 42nd Civil Chamber of the Munich Regional Court I largely upheld claims by the collecting society GEMA against the US AI music generator Suno, in case 42 O 763/25. The court found infringement of six well-known compositions, including "Forever Young" and "Mambo No. 5". Its press release says the judgment is not final. Commentators describe it as the first European ruling to make training that took place entirely in the United States actionable before a German court.
The strongest case for GEMA
GEMA's position deserves a fair statement. Composers and lyricists earn a living from licensing, and a model that can emit a recognisable "Atemlos durch die Nacht" or "Rasputin" from a simple prompt is, on its face, a substitute for the licensed work. GEMA's leadership argues that such systems store nearly complete works at scale. If a company could avoid liability by running its training job in a friendlier jurisdiction, EU copyright law would be easy to sidestep. GEMA also warns that companies could relocate to copyright-unfriendly jurisdictions unless European lawmakers step in.
What the court actually decided
According to the court's press release and legal commentary, the chamber treated four acts as infringing: copying for training in the US, reproduction through memorisation inside the model, offering the model, and generating outputs. The court found that training copies were made in the US, allegedly by ripping audio from YouTube while circumventing technical protections, and it assessed that conduct under US law. It rejected Suno's fair use defence under 17 U.S.C. § 107. It distinguished US decisions in which training data was not substantially recoverable from outputs.
For Germany, the court applied § 16 of the Urheberrechtsgesetz (UrhG) to find that memorising works in the model was reproduction. It held that the text-and-data-mining exception in § 44b UrhG does not cover that memorisation. Bristows' analysis says the court relied on the protection-country principle and on model weights held on German servers, and that the injunction expressly reaches copying "within the territory of the United States".
Where the reasoning is sound
The memorisation and output findings follow the same logic as GEMA's first win, the 11 November 2025 judgment against OpenAI (case 42 O 14139/24). CMS's summary says that court held that storing lyrics in model parameters and reproducing them in outputs infringes, and that the TDM exception did not apply because actual lyrics, not abstract information, were transferred into the model. It also notes that the ruling addressed post-training uses rather than the training process itself.
That is a proportionate line. A model that regurgitates a specific song is doing something a licensing market can price. Liability tied to demonstrable reproduction is evidence-based: a court can test the output and see it. It does not treat learning statistical patterns from music as infringement in itself.
Where it overreaches
The Suno judgment is riskier on three fronts.
First, extraterritorial reach. An order directed at US training conduct means a regional court in one member state is regulating what happens on US servers, under a theory that turns on a German hook (weights stored in Germany). Other jurisdictions can copy the move. If every court applies its own law to every training run, developers face a patchwork of injunctions with no single compliance target. Small and open-source developers, who cannot afford parallel litigation in several countries, would bear the heaviest burden.
Second, the US law question. A German court ruling that US fair use does not apply is unusual, and US courts have gone in different directions. The Ninth Circuit's September 2026 decision in Doe v. GitHub, as EFF reports, rejected an expansive reading of DMCA § 1202. It held that a new work lacking copyright information is not, by itself, evidence that someone removed it. That decision concerns a different provision. It does show US courts declining to stretch copyright doctrines to fit AI output, while Munich moves in the opposite direction.
Third, the remedy may outrun the harm. The court's findings rest on six works and on outputs that were recognisably close to them. A remedy that bars training on the collecting society's repertoire more broadly, if it stands on appeal, could push developers to exclude European music from their datasets. That would reduce cultural diversity in generative tools without clearly increasing payments to creators.
What a proportionate rule would look like
The better path combines three elements. Liability should attach to memorisation and to outputs that reproduce protected works. That is where harm is measurable. Developers should be able to defend themselves with technical mitigation, such as output filters and deduplication, and with documented dataset provenance. And licensing should be easier to obtain: collective licences for AI training, offered on transparent terms, would give both sides something better than a lawsuit.
The judgment is not final, and higher courts, and eventually the Court of Justice of the EU, will have to decide whether territorial copyright can govern where a model is trained. Until then, developers serving German users should assume that memorised works in a deployed model are a live liability. Lawmakers should resist reading Munich as a licence to regulate global training practices court by court.