Landgericht München I (42nd Civil Chamber), Case No. 42 O 763/25, decided July 31, 2026, is the first European judgment to hold a generative-AI music tool liable for copyright infringement—and it does so by asserting jurisdiction over training conduct occurring in the United States, applying U.S. copyright law to that conduct, and rejecting the fair-use defense.
GEMA, the German collecting society, sued Suno Inc. on behalf of the composers of six well-known musical works. It was undisputed that Suno’s model had been trained on them. The chamber enjoined four categories of conduct: reproduction for training in the United States, reproduction through memorization within the model in Germany, communication to the public through the offering of the model, and reproduction and communication to the public through outputs.
Market substitution ran through the fair-use analysis—Suno bore the burden of disproving market harm, and the substantially similar outputs on its free tier could substitute for the originals.
The court also rejected Suno’s reliance on the §44b text-and-data-mining exception, reasoning that the memorized works were not reproduced for the purpose of TDM and that Suno lacked lawful access to the training material.
The memorization finding—treating protected works retained in model parameters as reproductions under §16 UrhG—contrasts sharply with the England and Wales High Court’s decision in Getty Images v. Stability AI, [2025] EWHC 2863 (Ch), which held that Stable Diffusion’s final model weights did not store or reproduce Getty’s copyrighted images.
The ruling extends the same chamber’s reasoning in GEMA v. OpenAI (November 2025) from a text-to-text tool to a music-generation context.
The judgment is first-instance and non-final. An appeal is expected, particularly on the extraterritorial training issue and the memorization standard. The broader TDM questions remain before the BGH in Kneschke v. LAION, I ZR 281/25, and the CJEU in Like Company v. Google, C-250/25.
Compared to U.S. law, this holding cuts against the approach in Bartz v. Anthropic and Kadrey v. Meta, where particular uses of copyrighted works to train LLMs were found to be fair use. But Kadrey did emphasize the potential importance of market dilution and of market-harm evidence, so the convergence appears to be on market effects: when AI outputs compete directly with protected works, the fair-use analysis may look very different from cases involving training alone. The pending U.S. litigation against Suno will provide another important test.
For developers serving EU users, the practical lesson: a U.S.-based training pipeline and a U.S. fair-use theory may not insulate an EU-facing service from European copyright liability.
