A German court has delivered one of Europe’s most significant decisions involving generative music, ruling that Suno infringed copyright through the unauthorized training, memorization and reproduction of six protected musical works.
The Munich Regional Court I largely upheld claims brought by the German collecting society GEMA on July 31, 2026. The judgment grants injunctive relief, access to information and entitlement to damages.
The ruling is not yet final. Suno disputes the decision and may appeal, while the amount of damages remains to be determined.
Key Facts
| Decision | July 31, 2026 |
|---|---|
| Court | Munich Regional Court I |
| Case | GEMA v. Suno Inc. |
| Case number | 42 O 763/25 |
| Main findings | Unauthorized training copies, model memorization and infringing outputs |
| Remedies | Injunction, disclosure of information and damages |
| Status | First-instance ruling, subject to appeal |
Six Well-Known Songs Were Examined
The case involved six compositions represented by GEMA:
- “Forever Young”;
- “Big in Japan”;
- “Mambo No. 5”;
- “Daddy Cool”;
- “Rasputin”;
- “Atemlos durch die Nacht.”
The dispute focused on protected musical elements such as melody, harmony and rhythm. Lyrics were not part of this particular case.
GEMA argued that the compositions had been copied during Suno’s training process, retained inside versions of its models and reproduced through outputs that remained recognizably similar to the original works.
The court largely accepted that argument.
The central issue was not simply whether Suno had analysed copyrighted music. The court examined whether the protected works remained inside the models in a form that could later be reproduced.
The Training Dataset Contained Protected Works
According to the court, Suno’s training catalogue included the six disputed compositions.
The judgment also states that the company used stream-ripping techniques to obtain music from YouTube while circumventing the platform’s Rolling Cipher, a technical measure intended to prevent unauthorized downloading.
This finding broadens the legal debate beyond the usual question of whether AI training can qualify as fair use or text and data mining.
The method used to acquire training material may also create liability. The origin of the files, the absence of licences and the circumvention of technical protections can all become relevant.
The Court Found Memorization Inside Suno’s Models
Suno argued that its models did not store copies of songs. According to the company, model parameters contained only mathematical patterns learned from training data.
The court rejected that explanation in relation to the six works examined.
It concluded that the compositions were reproducibly contained in Suno model versions v3.5 and v4. By comparing the original music with generated outputs, the judges found that protected material had been retained in a form that could later be extracted.
The complexity and duration of the similarities made coincidence unlikely, according to the ruling.
The chamber therefore treated the presence of the works inside the models as an infringement of the reproduction right under German copyright law.
Simple Prompts Produced Recognizable Results
The prompts used during the case were another decisive factor.
GEMA entered song titles, original lyrics and general stylistic instructions into Suno. The prompts did not contain the protected melodies, harmonies, rhythms or complete arrangements.
Suno argued that responsibility belonged primarily to the users who generated the outputs.
The court disagreed.
It considered the instructions relatively simple and open-ended. Because the users had not supplied the musical elements responsible for the resemblance, the court concluded that Suno’s models had substantially determined the final results.
The ruling suggests that an AI provider may remain responsible when simple prompts produce outputs that reproduce protected musical material stored by the model.
The Provider Can Be Held Responsible
Generative AI companies frequently describe themselves as neutral technology providers, arguing that users are responsible for the content they create.
The Munich court placed greater emphasis on Suno’s own role in selecting training material, operating the models and determining how memorized content influenced the generated music.
This does not mean that every similarity produced by every AI system will automatically create liability.
The decision concerns six specific compositions, particular Suno model versions and evidence examined in this individual case.
However, it gives rights holders a clearer legal argument: when protected music is retained and reproduced by a commercial system, responsibility may remain with the developer rather than shifting entirely to the user.
The Court Rejected Suno’s Fair Use Defence
Although Suno’s models were trained in the United States, the German court examined the legality of those activities and considered the company’s defence under US fair use principles.
The chamber concluded that fair use did not apply because substantial elements of the protected works appeared in the outputs.
The court distinguished the case from other recent AI lawsuits in which plaintiffs had not demonstrated that training works could reappear significantly in generated content.
In this case, simple prompts produced recognizable musical similarities. The court therefore found that the relevant fair use factors weighed against Suno.
The judgment does not bind courts in the United States, but it shows that European courts may examine foreign training activities when they are closely connected to alleged infringement in Europe.
Why This Decision Matters
The ruling addresses several of the music industry’s most important unresolved questions:
- Can protected songs be copied without permission to train a commercial model?
- Can copyrighted music remain memorized inside AI parameters?
- Who is responsible when a model reproduces recognizable musical elements?
- Can fair use apply when original material reappears substantially in outputs?
- Does the method used to obtain training files affect liability?
- Can a European court examine training activities conducted abroad?
The court answered those questions strongly in favour of GEMA under the specific circumstances of the case.
Damages Have Not Yet Been Determined
The judgment recognizes GEMA’s entitlement to damages, but it does not announce a final financial award.
Suno may be required to disclose information needed to calculate the amount if the ruling becomes enforceable.
The final figure will depend on additional evidence, the method used to calculate losses or licensing value and the outcome of any appeal.
Reports describing Suno as already having paid a confirmed penalty would therefore be inaccurate.
Suno May Appeal
Suno has rejected the court’s conclusions and is considering its legal options.
An appeal could challenge the findings on jurisdiction, training, memorization, user responsibility, text and data mining or the similarity between the generated outputs and the original works.
For that reason, the judgment should be described as a major first-instance decision rather than a final resolution of European AI music law.
What the Ruling Could Change
If the reasoning survives appeal, generative music companies may face stronger pressure to:
- document the origin of their training datasets;
- obtain licences before using protected catalogues;
- prevent models from memorizing complete works;
- filter outputs that reproduce recognizable compositions;
- publish clearer information about training practices;
- create effective complaint and removal procedures.
The judgment also strengthens the argument that model training and generated outputs cannot always be treated as separate legal stages.
When a protected work is copied during training, retained by the system and reproduced through a simple prompt, the full process may be examined as one connected chain.
A Turning Point for AI Music Licensing
GEMA’s case is not an attempt to prohibit every use of artificial intelligence in music.
Its central demand is that composers, publishers and rights holders should be compensated when their works are used to build commercial generative systems.
The Munich judgment supports that position by rejecting the idea that converting music into model parameters automatically removes copyright responsibility.
Suno can still appeal, and other courts may reach different conclusions in future cases. The legal framework surrounding generative music remains unsettled.
Nevertheless, the ruling sends a clear warning to AI developers.
When copyrighted music remains memorized inside a model and returns through recognizable outputs, the machine may still be reproducing protected human creativity.
