A Narrow Case With a Wide Reach
On July 31, 2026, the 42nd Civil Chamber of the Munich Regional Court I ruled against Suno Inc. in GEMA v. Suno (case no. 42 O 763/25), the first European judgment against a generative AI music tool. The court found Suno liable for four distinct acts tied to six well-known songs — including "Atemlos," "Rasputin," "Mambo No. 5" and "Forever Young" — reproduction for training purposes carried out in the United States, memorisation and reproduction inside its model in Germany, communicating the model to the public, and reproducing the works again through its outputs. The court ordered Suno to disclose the scale of its use of GEMA's catalogue and held it liable for damages, with the amount to be determined later. The judgment is not yet final, and Suno has said it will "explore all options, including an appeal."
What makes the ruling significant isn't the finding of infringement itself — courts routinely hold companies liable for reproducing protected works without a licence. It's where the infringement was found to occur. Suno's training happened on US servers, using US-hosted data, and would ordinarily fall under American copyright law and US courts. The Munich court instead relied on Germany's Collecting Societies Act (VGG) — a statute written to give collecting societies like GEMA a convenient home venue for royalty disputes — to assert jurisdiction over conduct that took place entirely outside the EU, then applied US copyright law to that conduct and rejected Suno's fair-use defense.
The Case for the Ruling
GEMA's underlying complaint deserves to be taken seriously on its own terms. The court's press release describes Suno's system as capable of reproducing substantially similar versions of the original recordings — GEMA's general counsel, Kai Welp, said the models "apparently store nearly complete works in substantial quantities." Separate reporting on the case describes Suno using stream-ripping techniques against YouTube's "rolling cipher" protection to extract the source audio in the first place — not scraping public text, but circumventing a technical measure to obtain protected recordings. If a free tier of a commercial product outputs songs that are functionally interchangeable with the originals, GEMA's point that this substitutes for licensed use, rather than merely referencing it, is a fair one. Rightsholders who license their catalogues to streaming services and film studios have a legitimate claim that an AI company should not be exempt from the same bargain just because its output is machine-generated. Enforcement failures here would tell every AI lab that unlicensed, near-verbatim reproduction is cost-free as long as a large-enough model launders it.
Where the Reasoning Overreaches
The problem is the vehicle, not the verdict. Using a venue statute designed for domestic royalty administration to reach US-based training conduct and apply US law to it is a novel and aggressive jurisdictional move — one that, if it holds up on appeal, invites every jurisdiction with a national champion collecting society to claim the same authority over conduct that happens entirely abroad. That cuts both ways: courts in jurisdictions with far weaker speech and due-process protections than Germany's could cite this precedent to assert authority over content moderation or AI outputs generated in the US or EU. A fragmented patchwork of national courts each asserting extraterritorial reach over global AI training pipelines is a worse outcome for the open internet than a single, predictable rule — even an imperfect one.
It also matters that this case is narrower than the headlines suggest. The court's finding rests on memorisation and near-identical reproduction of specific recordings obtained by circumventing a technical protection measure — not a general ruling that training on copyrighted material is unlawful. The EU's own Copyright in the Digital Single Market Directive already provides a working framework: Article 4 permits commercial text-and-data mining unless a rightsholder opts out "in an appropriate manner, including machine-readable means" (EUR-Lex). That regime, alongside the Munich chamber's own November 2025 ruling against OpenAI establishing that memorised, reproducible outputs fall outside the TDM exception, already gives AI developers a defined line: mine broadly, but don't ship a product that regurgitates the training data. Suno's case looks like a company that crossed that line by extracting protected recordings through DRM circumvention — not a referendum on AI training generally.
What Should Happen Next
The sensible response is not to treat this as license to expand extraterritorial copyright enforcement by analogy, but to press EU lawmakers to clarify the boundary between legitimate mining and unlicensed reproduction in statute, rather than leaving it to one regional chamber's reading of a collecting-societies venue rule. Predictable, EU-wide rules — not a rolling series of national court experiments — are what let both AI developers and rightsholders plan. GEMA's underlying grievance, that a commercial product substituted for licensed music without paying for it, is worth vindicating. The jurisdictional theory used to get there is worth appealing.