On 31 July 2026, the Munich Regional Court ruled that Suno, the AI music generator, infringed copyright in six musical works represented by the German collecting society GEMA. The suno copyright ruling is the first European court decision holding an AI music generator liable, and it rejected Suno's argument that its U.S.-based training was protected by fair use. The judgment is not final, and Suno has said it is weighing an appeal.
TL;DR
- Munich Regional Court (Landgericht München I), 42nd Civil Chamber, case 42 O 763/25, decided 31 July 2026. Filed 21 January 2025. Last verified: 1 August 2026.
- The court found Suno's v3.5 and v4 models had memorised protected works, rather than storing only abstract statistical patterns.
- Six works were at issue, including "Atemlos durch die Nacht", "Forever Young", "Big in Japan", "Rasputin", "Daddy Cool" and the refrain of "Mambo No. 5".
- The U.S. fair use defence under 17 U.S.C. § 107 failed because the protected works were substantially present in the outputs, unlike in the Bartz and Kadrey decisions.
- Suno, not the user, was held responsible, because prompts supplied only lyrics and a style, with no melody, harmony, rhythm or arrangement.
- Orders: cease the infringement, disclose revenue data, pay damages. The damages figure has not been set.
What exactly did the Munich court decide?
The 42nd Civil Chamber is Munich's copyright-specialised chamber. It handled the case as 42 O 763/25 after GEMA filed in January 2025. The court's reasoning runs along three linked findings.
First, storage. Suno's v3.5 and v4 models were held on servers located in Germany. Because the court concluded that the protected works were embedded in the model parameters, keeping those models in Germany amounted to reproduction under § 16 of the German Copyright Act (UrhG). The text and data mining exception in § 44b UrhG did not cover it.
Second, availability. Offering the model to users for generation was treated as an unauthorised act of communication to the public under § 15(2) UrhG.
Third, output. The generated tracks contained recognisable original elements of the six works, which is what pulled the whole chain together. The court's press release is published by the Bavarian Ministry of Justice and remains the authoritative account of the decision (justiz.bayern.de).
Why did the fair use defence fail?
Suno argued that the training copies were made in the United States and should therefore be assessed under U.S. law. The court accepted the premise. Applying the territoriality principle, it agreed that U.S. law governs copies made on U.S. soil, then examined 17 U.S.C. § 107 on its own terms and concluded the training was not fair use.
The distinguishing factor was the output. In the U.S. decisions in Bartz and Kadrey, the training material was not delivered back to users in recognisable form. Here, the court found the opposite: substantial portions of the protected compositions surfaced in what Suno's users generated. Once that link is established, the "transformative use" argument that carries most of the weight in a fair use analysis becomes much harder to sustain.
That is a narrower holding than the headlines suggest. The court did not declare that AI training is unlawful everywhere. It said that when the training material is reproducible from the model, the training and the output cannot be assessed as separate, unrelated events.
What does the "memorisation" finding change?
This is the part with the widest reach. A standard defence in generative AI litigation is that a model learns statistical relationships and retains no copies. The Munich court examined that claim and did not accept it for Suno's models, finding instead that the content of the training data was present in the parameters.
The practical consequence is evidentiary. If a claimant can demonstrate that a model reproduces protected material on ordinary prompts, the "only mathematics" argument stops being a shield and becomes a contested factual question the defendant has to win. That applies to text, image and code models as much as to music, which is why the ruling is being read well outside the music sector. It also connects to a broader pattern of law arriving after deployment rather than before it, something we examined in the context of facial recognition at protests.
The court also noted that Suno acknowledged using stream-ripping to extract works from YouTube, which involved circumventing YouTube's Rolling Cipher protection measure. Technical protection circumvention is a separate legal exposure in most European jurisdictions and does not depend on how the model works internally.
Who is liable when a user writes the prompt?
Suno's position placed responsibility with the person typing the prompt. The court rejected that, and the reasoning is specific rather than sweeping.
The prompts in evidence contained lyrics and a description of musical style. They did not specify melody, harmony, rhythm or arrangement, which are the elements that made the outputs recognisable. Since the user did not supply those elements, the court attributed the creative overlap with the protected works to the system rather than to the user.
This leaves an open question that future cases will have to answer: a prompt engineered in detail to reproduce a specific work would shift the analysis. For now, the default allocation of liability in Germany sits with the operator, not the subscriber.
How does this fit Suno's other legal exposure?
The Munich case is one front among several. Suno raised over $400 million in a Series D in June 2026 at a $5.4 billion valuation, and reports more than 2 million paying subscribers against over 100 million total users. That scale is what makes the litigation commercially material.
| Matter | Status |
|---|---|
| GEMA v Suno (Munich, 42 O 763/25) | Infringement found, 31 July 2026; appeal possible |
| RIAA action for UMG, Sony, Warner (U.S., June 2024) | Warner settled November 2025 with a licensing deal; UMG and Sony continue |
| Koda (Denmark) | Separate European proceedings ongoing |
| Artist class actions vs Suno and Udio | Supported by over 1,800 artists |
There is also a direct precedent. In November 2025, the same Munich court found for GEMA against OpenAI over ChatGPT's reproduction of German song lyrics. The Suno decision goes further, because it covers complete compositions rather than text alone.
GEMA represents roughly 95,000 members in Germany and more than 2 million rightsholders worldwide. Its chief executive, Tobias Holzmüller, framed the outcome as a call for licensing rather than prohibition, arguing the generative AI market has operated without basic transparency and fairness. The Warner settlement suggests licensing is a workable route when the commercial incentives line up.
What should AI builders do differently now?
Concrete steps, in rough order of urgency:
- Test your own models for regurgitation. Run adversarial prompts against known training material. If protected content comes back recognisably, you have the exposure the Munich court identified.
- Map where model weights are stored. The German reproduction finding turned on servers in Germany. Hosting geography is now a legal variable, not just a latency one.
- Check EU AI Act readiness. The Act's copyright transparency and compliance obligations become enforceable on 2 August 2026, with machine-readable marking deadlines for existing systems by 2 December.
- Document data provenance. Circumvention of technical protection measures, such as stream-ripping, creates liability that is independent of any training-exception argument.
- Price licensing into your model. Collecting societies are the practical counterparty for music. Budget for it the way you budget for inference costs.
The valuation-versus-liability tension here is not unique to Suno. It appears whenever capital moves faster than the legal groundwork, a dynamic worth comparing with how AI labs are being valued and with the pricing pressure now visible in frontier model economics. Teams evaluating open-weight alternatives face the same provenance questions, since open weights do not resolve what the weights were trained on.
Frequently asked questions
Q: Is the Suno copyright ruling final? A: No. The court's decision of 31 July 2026 can be appealed, and Suno has said it disagrees and is evaluating options.
Q: Does this mean AI music generation is illegal in Germany? A: No. The court addressed one operator, six works and specific model versions. The finding depended on the works being reproducible from the model and recognisable in the outputs.
Q: Which songs were involved? A: Six works: "Atemlos durch die Nacht", "Rasputin", "Big in Japan", "Forever Young", "Daddy Cool", and the refrain of "Mambo No. 5 (A Little Bit of...)".
Q: Can a German court rule on training that happened in the United States? A: Yes, under the approach taken here. The court applied territoriality — assessing U.S. training copies under U.S. fair use, while assessing German server storage and availability under German law.
Q: How much does Suno have to pay? A: The amount is not yet determined. The court ordered Suno to stop the infringement and to disclose revenue information, which is the usual precursor to quantifying damages.
Q: Does the EU AI Act change any of this? A: It adds a separate regulatory layer. Copyright transparency and compliance obligations become enforceable from 2 August 2026, with machine-readable marking required for existing systems by 2 December.
Last verified: 1 August 2026. Primary source: Munich Regional Court press release, published by the Bavarian State Ministry of Justice (case 42 O 763/25). Damages figures and appeal status were unresolved at the time of writing.
Corrections log: No corrections issued.
This article was researched and written with AI assistance under human editorial review. See how we work.
Discussion
0 comments