GEMA v Suno: German court extends AI copyright reach

On 31 July 2026, the Munich Regional Court I (Landgericht München I) ruled largely in favour of GEMA, the German music collecting society, against US AI music generator Suno (case no. 42 O 763/25). It is the second win in under a year for GEMA against a generative AI provider (see our commentary on the GEMA v OpenAI ruling of last November here, currently under appeal), and the first European ruling to hold that AI training conducted entirely outside the EU can still cause German copyright liability. In this article, we discuss the GEMA v Suno decision and its implications, in particular with regard to questions of jurisdiction, memorisation, the interplay with fair use under US law, and extraterritorial remedies.

Background

In its claim filed on 21 January 2025, GEMA sought an injunction, information (revenue disclosure) and a declaration of entitlement to damages on behalf of the composers of six well-known works including "Rasputin" and "Daddy Cool" (Frank Farian), "Big in Japan" and "Forever Young" (Alphaville/Marian Gold et al.), and the refrain of "Mambo No. 5" (Lou Bega).

Unlike in GEMA v OpenAI, it is not the song lyrics but the musical compositions of the songs that are at issue here. Suno accepted that its training dataset included the works and that they were extracted via “stream-ripping” from YouTube. Stream-ripping is a method by which media streams are saved as permanent, downloadable files on a personal device. It circumvents YouTube’s "rolling cipher" technical protection measure. GEMA's evidence consisted of outputs generated from simple prompts naming the title, lyrics and musical style, which it said were "confusingly similar" to the originals.

Arguments

Suno raised several defences, claiming that: (i) the works were not shown to be protected, or recognisable in the outputs; (ii) a model's weights reflect only generalised statistical patterns rather than the training data itself; (iii) any similarity resulted from GEMA's own complex, iterative prompting, breaking the chain of attribution; (iv) US training was covered by fair use, a question the court in any event lacked jurisdiction to decide; and (v) so far as German law applied, the text and data mining (TDM) exception under section 44b of the German Copyright Act (Urheberrechtsgesetz, UrhG) applied.

Decision

Jurisdiction. The Munich court accepted jurisdiction over the US training conduct itself, relying on section 131(1) and (2) of the German Collecting Societies Act (Verwertungsgesellschaftengesetz, VGG), which it read as giving collecting societies (but not other rightsholders, such as publishers or labels) a special forum of factual connection extending to related infringements abroad.

Memorisation. The court found the works were "memorised", i.e. reproducibly contained, in versions of Suno's models (reportedly v3.5 and v4) stored on German servers. This was evidenced by their substantial reconstruction from short, open-ended prompts. This storage itself infringed the reproduction right (section 16 UrhG) and, per the court's earlier GEMA v OpenAI reasoning, fell outside the TDM exception, which only covers reproductions genuinely necessary for analysis, not full retention of works in model parameters.

The court drew on the Hamburg court’s landmark ruling on TDM exceptions in AI model training in Kneschke v LAION (see our commentary regarding the September 2024 ruling here, upheld in December 2025 by a higher court – case no. 5 U 104/24 [German language]), in which the court confirmed that dataset creation may fall within TDM exceptions, while taking a rightsholder-friendly view of what counts as a machine-readable opt-out.

Further, the CJEU is due to assess the issue of memorisation in a lawsuit brought by Hungarian publisher Like Company against Google (C-250/25) regarding alleged copyright infringement via Gemini’s training and outputs, and the application of TDM exceptions. The Advocate General’s opinion is expected in September 2026.

Outputs and attribution. Suno, not its users, was held responsible for the outputs, which remained recognisable reproductions and unauthorised communications to the public (section 15(2) UrhG). The court rejected Suno’s "complex prompting" defence: the prompts (repeated over 100 times in places) were simple and open-ended, and it was the model's architecture and memorised training data, not the prompt, that substantively determined the output.

Fair use. Applying the territoriality principle (Schutzlandprinzip), the court conducted its own analysis of US fair use under 17 U.S.C. § 107 for the US training conduct, rather than declining jurisdiction or deferring to expert evidence. It found fair use inapplicable, and expressly distinguished Bartz v Anthropic and Kadrey v Meta. In those cases outputs were not shown to closely reproduce training works and no market harm was established, whereas here outputs were substantially similar, from non-specific prompts, and the court considered Suno's service a substitutable competitor to the original works.

EU AI Act. The court held that compliance with Article 53(1)(c) and (d) GPAI transparency duties does not itself discharge copyright liability (relying on Recital 107). It treated the Act as intended to ease enforcement for rightsholders, not to displace copyright law.

Remedies and appeal. The court granted injunctions prohibiting Suno from (i) using the relevant works to train its models, (ii) storing the works within its models, (iii) offering the model trained on the works in Germany, and (iv) creating infringing adaptations through the outputs. Notably, the judgment explicitly ordered Suno to stop copying the works “within the territory of the United States of America for the purpose of training an artificial intelligence (AI) model to generate music” (see section K1.6(aa) of the judgment [our translation]), thereby giving the injunction extra-territorial effect. Damages were not quantified. The judgment is not final and may be appealed.

Wider implications

The Munich court’s decision has arrived at a moment when courts and policymakers on both sides of the Atlantic are grappling with the intersection of generative AI and copyright law.

Suno is currently defending lawsuits for alleged copyright infringement brought by major record companies including Universal Music Group (UMG) and Sony in the US. These cases are expected to test the US fair use doctrine in relation to AI model training.

In November 2025, Suno settled a lawsuit by Warner Music and reached a licensing deal. However, in June of this year, the American Federation of Musicians sued Warner (and UMG with regard to a licensing deal it struck with Udio, another GenAI music production platform), alleging that the labels’ AI licensing deals fail to compensate or inform musicians whose performances were used in sound recordings licensed for AI training.

Similarly, in the case of GEMA, litigation and licensing are running in parallel: just over a week before the Munich court’s judgment in July 2026, it launched PLAI by GEMA, which it describes as Europe’s first fully licensed, scalable, music dataset of its kind. It allows AI music tool providers to customise rights portfolios from around 178,000 sound files across more than 60 genres, and aims to “create a sustainable basis for collaboration between the high-tech and creative sectors”.

Practical takeaways

For AI providers, the judgment reinforces that offshoring training will not, on this reasoning, insulate against a (German) collecting society's claims, and EU AI Act compliance measures should not be treated as a copyright defence.

For rightsholders, the case suggests that unlicensed training on protected repertoire carries real litigation exposure, and that German courts are willing to find a model's ability to reproduce protected content on simple prompting to be evidence of unlawful reproduction, a theory TDM defences have not yet displaced at first instance.

Until the OpenAI appeal and any appeal in this case are resolved, and the CJEU provides guidance in Like Company, the practical position remains as after the OpenAI ruling: document lawful data access, take machine-readable rights reservations seriously, and treat memorisation as a live commercial risk.