What the court decided
On July 31, 2026, the 42nd Civil Chamber of the Landgericht München I ruled that Suno Inc., the US-based text-to-music generator, infringed the copyrights of six musical works in GEMA's repertoire — including "Atemlos durch die Nacht," "Daddy Cool," "Rasputin," "Big in Japan," "Forever Young," and the refrain of "Mambo No. 5" (Landgericht München I press release, case 42 O 763/25). The court granted GEMA injunctive relief, ordered Suno to disclose revenue tied to the infringement, and found the company liable for damages, with the amount to be determined separately (GEMA press release, July 31, 2026).
What makes the ruling consequential isn't just the outcome — it's the reach. The court found that Suno's training, which occurred in the United States, was still actionable under German law because GEMA, as a collecting society, has a special jurisdictional basis to sue in Germany over infringements with a sufficient connection to German rightsholders. It then applied US fair-use doctrine to the American training conduct and found against Suno anyway, distinguishing the case from Bartz v. Anthropic and Kadrey v. Meta on the grounds that Suno's outputs were "substantially similar" to the protected works and generated from generic, non-targeted prompts (ReedSmith legal analysis). Separately, the court held that the EU's text-and-data-mining exception under Section 44b of the German Copyright Act — the domestic implementation of Article 4 of the EU's 2019 DSM Directive — does not cover "memorization," where a model retains and can reproduce a work's original elements rather than merely learning statistical patterns from it.
The case for the ruling
GEMA's underlying claim deserves to be taken seriously on its own terms. Songwriters and composers have spent decades building a collective-licensing system specifically so that commercial users of music — broadcasters, streaming platforms, venues — pay for what they use. Suno is, by GEMA's account, a commercial subscription product that reproduces recognizable musical elements of copyrighted songs, generating revenue while creators of the underlying works see none of it. The court found that Suno's system used stream-ripping-derived training data and that six well-known songs could be reproduced in near-identical form from simple prompts (JUVE Patent). If a human producer sampled "Rasputin" without a license and sold the result, no one would call that fair use. A court asking whether an AI company should get a pass merely because the copying happens inside a matrix multiplication, rather than a sampler, is asking a fair question, and GEMA's win is the second in less than a year following its November 2025 case against OpenAI.
Why the jurisdictional reach is the real problem
The infringement finding on outputs that closely mimic specific, identifiable songs is defensible and consistent with how copyright has long treated close copying, regardless of the tool used. But the court's jurisdictional theory — that training conducted wholly outside Germany, by a company with no German infrastructure for that training, can still trigger German copyright liability because a German collecting society is the plaintiff — is a significant expansion. Taken to its logical end, it would let any EU member state's judiciary regulate how AI models are trained anywhere in the world, so long as a local rights-holder can be found to sue. That is a recipe for models being trained to the most restrictive available jurisdiction's standard, not the EU's own harmonized one, since the DSM Directive's TDM exception was specifically designed to give EU-based AI development legal certainty that this ruling now undercuts for foreign developers reaching EU repertoires.
The ruling also blurs a line that copyright law needs to keep sharp: the difference between training on a work and reproducing that work's protected expression in an output. Conflating the two — treating any "memorization" detectable in a model's weights as equivalent to actionable reproduction — risks capturing ordinary statistical learning that never surfaces in a user-facing output. The court says memorization plus reproducible outputs is what mattered here, which is narrower than a blanket "training itself infringes" rule. But the reasoning, if adopted more broadly, gives future plaintiffs an incentive to allege memorization first and litigate the distinction later, chilling investment in EU-facing AI products long before any output ever reproduces a specific song.
What happens next
Suno can appeal to the Munich Court of Appeal (Oberlandesgericht München), and given the stakes for the jurisdictional theory, an appeal looks likely. Notably, Warner Music — after suing Suno alongside Universal and Sony — struck a licensing partnership with the company in late 2025, suggesting commercial licensing markets are already forming as an alternative to scorched-earth litigation (JUVE Patent). That is the outcome regulators and courts should want to encourage: rightsholders get paid, AI developers get legal certainty, and neither side needs a court to invent extraterritorial jurisdiction to get there. The Munich court's narrow finding — that verbatim-adjacent outputs of identifiable songs infringe — is sound. Its jurisdictional reasoning is the part Brussels and Berlin should be watching closely before it hardens into EU-wide practice by default.