A Munich court ruled on 31 July 2026 that Suno infringed German copyright by training its AI music model on protected songs — on reasoning specific to German law, not a distinction EU legislation itself draws.
GEMA, the German collecting society for musical performance and mechanical reproduction rights, filed suit against Suno at the Landgericht München I (Munich Regional Court I) in January 2025, case number 42 O 763/25. The 42nd Civil Chamber, which specialises in copyright law under presiding judge Elke Schwager, heard the case in March 2026. Judgment was originally scheduled for 12 June but was postponed to 31 July, when the court ruled in GEMA's favour.
GEMA's claim centred on six well-known compositions: "Rasputin" (Frank Farian and co-writers) and "Daddy Cool" (Frank Farian), "Forever Young" and "Big in Japan" (Alphaville), "Atemlos" (Kristina Bach, recorded by Helene Fischer), and "Mambo No. 5" (Lou Bega). GEMA demonstrated that entering simple prompts into Suno produced audio outputs that corresponded, in melody, harmony, and rhythm, to these six works. The court found this undisputed: Suno had trained its models on the songs, extracted using stream-ripping techniques that circumvented YouTube's technical protection measures, and the resulting outputs were too close to the originals to be coincidental.
The court ordered Suno to stop reproducing the works and to stop using them for training, to disclose revenue connected to the infringement, and to pay damages, with the amount still to be determined. The judgment is provisionally enforceable pending appeal, and Suno has said it will consider all available options, including an appeal.
The evidential method mattered as much as the legal argument. GEMA's case did not rest on showing that the six songs were somewhere in Suno's training data in the abstract. It rested on demonstrating, through repeated prompting, that the system could reliably reproduce recognisable elements of specific, named works on demand. GEMA entered simple prompts describing a musical style and received outputs it says corresponded closely to the originals in melody, harmony, and rhythm across all six songs.
This method mirrors the approach GEMA used successfully against OpenAI eight months earlier, where the court accepted that entering prompts such as asking for the text of a named song, and receiving lyrics that were largely unchanged from the original, was sufficient to establish that the lyrics were fixed in the model's parameters.
In that earlier judgment, the court drew an analogy to lossy compression: just as an MP3 file does not store a perfect copy of a recording but retains enough statistical information to reconstruct a recognisable version, the court held it was sufficient that a model could generate statistically probable outputs that recognisably reproduced the protected work. The court rejected OpenAI's argument that GEMA needed to identify specific, definable data inside the model itself. What mattered was the output the system was capable of producing, not the internal representation.
Suno's central defence rested on Germany's text and data mining exception, Section 44b of the Urheberrechtsgesetz (UrhG), which implements Article 4 of the EU's Digital Single Market Directive. Section 44b permits reproductions of lawfully accessible works for the automated analysis of text and data to extract information such as patterns, trends, and correlations. It is the provision that, in principle, makes AI training on copyrighted material lawful in Germany without a licence.
The court rejected the defence on a specific factual finding: Suno's models had not merely analysed the six songs to extract statistical patterns. They had retained the songs in a form that could later be reproduced, a process German courts describe as memorisation. Where an AI model can output content that reproduces the specific expression of a protected work, rather than only general patterns learned across a dataset, the reproduction falls outside Section 44b because the exception was never intended to permit that outcome. The court found the similarities too extensive and specific to be explained by coincidence, which meant memorisation, not mere analysis, had taken place.
The claims themselves rested on specific statutory rights. Reproduction is protected under Section 16 UrhG and the right of communication to the public under Section 15(2) UrhG. GEMA v. Suno found Suno liable on two separate grounds: storing the protected songs inside the model infringed the reproduction right, and serving outputs built on that stored material to users in Germany infringed the right of communication to the public.
The same statutory framework — the reproduction right, the communication right, and the text and data mining exception in Section 44b — formed the backbone of the earlier GEMA v. OpenAI judgment (case number 42 O 14139/24, 11 November 2025), which concerned song lyrics reproduced by ChatGPT rather than musical compositions generated by Suno.
It matters that GEMA v. Suno is a decision of the Landgericht München I, a first-instance regional court, interpreting a German statute. Section 44b UrhG is Germany's national implementation of an EU directive provision that gives member states latitude in how they transpose it, and German courts are not bound by a single, harmonised EU interpretation of where the line between analysis and memorisation sits. The Court of Justice of the European Union has not ruled on this question, and until it does, each member state's courts are free to draw the line differently under their own implementing legislation.
Even within Germany, the line is being tested rather than settled. In December 2025, the Hanseatic Higher Regional Court in Hamburg reached a different outcome on a related question. In Kneschke v. LAION, a photographer challenged a nonprofit's use of his image to build a dataset for AI research.
The Hamburg court found that Section 44b did apply, because LAION had reproduced the image only to extract correlations between the image and its text description for dataset construction, not to enable an AI model to later reproduce the photograph itself. The OLG Hamburg dismissed the photographer's appeal, confirming that Section 44b applied to LAION's use — though only a ruling from the Bundesgerichtshof, Germany's Federal Court of Justice, would establish a binding interpretation across all German regional courts.
The distinction the two lines of cases draw is coherent: LAION's use stayed at the level of analysis, while Suno's and OpenAI's models retained and reproduced the protected expression itself. But it is a distinction being worked out case by case, in first-instance and intermediate appellate courts, using a piece of German legislation. It is not a rule any court outside Germany is required to follow, and it says nothing about how courts elsewhere in the EU would apply their own national implementation of the same directive provision to the same facts.
A further feature of GEMA v. Suno has limited precedent outside this specific German procedural context. Suno's training took place in the United States, and ordinarily that would place the training conduct outside a German court's jurisdiction.
The Munich court instead relied on Section 131 of the Verwertungsgesellschaftengesetz (VGG), Germany's Collecting Societies Act, which grants collecting societies a privileged venue for claims connected to their statutory role and permits them to bundle related claims against a single infringer. This forum is available to collecting societies specifically; it is not a general rule that any rights holder can invoke against conduct that occurred abroad.
Having established jurisdiction, the court applied US copyright law to the training conduct that took place in the US, and German law to the reproduction and communication to the public that occurred once the model was hosted on servers in Germany.
On the US side, the court distinguished Suno's position from recent US rulings where courts found AI training fair use in part because the training materials were not reproduced for users in the outputs. GEMA v. Suno found the opposite pattern on the evidence: simple prompts produced outputs substantially similar to the original works, which meant the fair use defence did not survive contact with what the outputs contained.
The injunction itself reflects this dual scope. The operative part of the judgment orders Suno to stop reproducing the six works for training purposes in the United States — not only to stop offering the model in Germany. A German regional court has issued injunctive relief that reaches directly into conduct on US soil, a consequence of the collecting society jurisdiction route that any rights holder operating within the VGG framework can potentially invoke.
For rights holders, the practical significance of GEMA v. Suno is narrower than the headlines suggest. It does not establish that AI training on copyrighted music is unlawful across the EU, and it does not bind courts outside Germany, or even automatically bind other German regional courts on the same question.
What it establishes, within German law, is that a rights holder alleging memorisation needs to show the AI system can reproduce the specific expression of a protected work, not merely that the work was somewhere in the training data. GEMA built its case on that exact evidence: prompting the system and demonstrating that the outputs matched the originals too closely for coincidence.
For AI companies operating or serving users in Germany, the ruling raises the practical cost of relying on the text and data mining exception without also testing whether a released model can be prompted into reproducing specific protected works. A model that only ever produces generic, stylistically similar output has a stronger Section 44b defence than one that can be made to reproduce a named song on request.
For German rights holders and their representative societies, the ruling confirms that a well-documented pattern of matching outputs, gathered through systematic prompting and comparison against the original works, is treated by the Munich court as sufficient evidence of memorisation without requiring the claimant to show what the model's internal parameters contain.
GEMA has paired the litigation with a commercial alternative. On 23 July 2026, a week ahead of the Suno judgment, it launched PLAI by GEMA, a licensed dataset of roughly 178,000 audio files drawn from more than 60 genres, intended for AI developers who want to train on cleared repertoire. The timing signals the strategy directly: establish in court that unlicensed training on memorised works infringes, then offer a licensed alternative in parallel.
Both GEMA v. OpenAI and GEMA v. Suno remain open to appeal, and neither has been tested by the Bundesgerichtshof or referred to the Court of Justice of the European Union. Until an appellate court rules, or the CJEU is asked to interpret Article 4 of the DSM Directive directly, GEMA v. Suno is what it is: a first-instance German court's answer to a fact-specific question about one AI company's models, decided under German legislation that other EU member states have not applied the same way.
For AI companies operating in Germany, and for rights holders whose works are used to train them, that answer already carries real consequences. For the wider question of whether AI training infringes copyright across Europe, it is one data point, not a continental rule.
This post provides general information about a recent German court ruling and does not constitute legal advice. For advice on a specific matter, consult a qualified lawyer in the relevant jurisdiction.
Related Reading
Munich Regional Court stops Suno using GEMA-protected music — JUVE Patent
GEMA v. OpenAI – Judgment of the Regional Court of Munich I — Preu Bohlig
James Snell is the founder of Provlyn, a platform providing cryptographic prior proof of IP ownership. provlyn.com