The Munich Regional Court largely upheld GEMA's claims against Suno for injunctive relief, information and damages over six memorised musical works reproduced in the model and in its outputs

What was decided

On 31 July 2026 the 42nd Civil Chamber of the Landgericht Muenchen I largely granted GEMA's claims against Suno, the provider of an AI music generator, for injunctive relief, information and damages.

Six works were at issue: Atemlos durch die Nacht by Kristina Bach, Rasputin by Frank Farian, Fred Jay and George Reyam, Big in Japan and Forever Young by Marian Gold, Bernhard Lloyd and Frank Mertens, the refrain of Mambo No. 5 A little bit of by David Lubega and Christian Pletschacher, and Daddy Cool by Frank Farian. Infringements through the lyrics were not part of the case.

The chamber held that the claims succeed both on the reproductions made during training in the United States and on the reproduction in the model in Germany together with the rendering in the outputs.

The memorisation finding

GEMA's case was that the works were copied during training and memorised in the underlying model, and that output generation produced further infringements. When generating the disputed outputs it entered the original lyrics, the desired musical style and the work title. The prompts contained no instructions as to melody, harmony, rhythm or arrangement.

Suno's answer was that the training data is neither contained nor stored in the model, that the weights and parameters represent mathematically learned patterns and generalised features such as syntactic, semantic and contextual relationships, and that any similarity in the outputs comes from prompt-induced narrowing of the search space and statistically learned patterns.

The chamber was not persuaded. It found the works reproducibly present in the defendant's model versions v3.5 and v4, which were stored on servers in Germany. It described memorisation as the case where a model does not merely extract information from the training set but where the parameters specified after training also contain a taking over of the content of the training data, and it established that by comparing the works in the training data with the renderings in the outputs. Given the complexity and length of the pieces, it ruled out chance as the cause.

Memorisation therefore infringed the reproduction right under section 16 UrhG, and the chamber held that reproduction in the models is not covered by the text and data mining exception in section 44b UrhG.

Who is responsible for the output

Suno argued that the outputs resulted from complex, iteratively refined prompts by the claimant, so that the chain of attribution was broken by an intervening, purposive user act.

The chamber rejected that. The outputs were generated by simple, open-ended prompts that specified only the lyrics and the musical style. The defendant operates the models, selected the works as training data, trained on them, and is responsible for the architecture of the models and the memorisation of the training data. On the court's reasoning the models therefore determined the content of the outputs, and the defendant rather than the user bears responsibility.

The chamber went further and held that merely offering the model and the application for generating music already breaches the unnamed right of communication to the public under section 15(2) UrhG.

The press release also records that Suno used stream ripping techniques to extract and copy the works from YouTube, circumventing the rolling cipher, a technical protection measure implemented by the platform to prevent downloading of audio and video content.

A German court deciding fair use

The chamber held it had international jurisdiction under section 131(1) and (2) VGG for the claims arising from infringing acts on US territory. It read that provision as covering both local and international jurisdiction and as providing a special forum of factual connection for privileged collecting societies whose purpose is the administration of copyright.

Under the country of protection principle it then applied US law to those acts. Because the works are found again in the outputs, it held the reproductions were not covered by fair use under 17 U.S.C. 107.

The distinction it drew is the part worth reading twice. The chamber said the facts differ significantly from those in Bartz and Kadrey, in which two US courts treated the training of AI models as covered by fair use. In those cases the training data was not, or not substantially, made accessible to the user in the outputs. Here, simple and open-ended prompts produced outputs substantially similar to the original works. Applying the Supreme Court's Warhol framework, the chamber said all the factors to be examined spoke against the defendant.

What to take from a non-final judgment

The court closed its own summary with a single line: the judgment is not legally final. Treat everything above as one chamber's reasoning, capable of being reversed.

Even so, the structure of the reasoning is portable. If a claimant can show that a work is extractable from a model with a plain prompt, the argument shifts from the contested ground of whether training is permitted onto the much older ground of whether a copy exists. Sections 44b UrhG and the fair use factors are both harder to run once a court accepts that the model holds a reproducible copy.

For providers, the operational implication is memorisation testing before release, and a record of it. For rightsholders, it is the prompt log: what the chamber credited was that the prompts were simple and open-ended, not that they were clever.

What we did not verify

We opened press release 16 of the Landgericht Muenchen I dated 31 July 2026 in full and took every fact and quotation from it. It is written by Cornelia Kallert, presiding judge at the Landgericht Munchen I, in her capacity as press spokesperson. The release does not say which chamber she presides over.

We did not read the judgment itself, which the press release does not publish, and we did not read the pleadings, the expert evidence on memorisation, or the Bartz and Kadrey decisions the chamber distinguished. We did not verify whether an appeal has been filed.

We refuse to state the damages figure, the exact scope of the injunction or which of GEMA's claims were not granted. The press release says the claims were largely upheld and does not quantify them, so we do not either.

Related briefings

• GEMA v OpenAI in Munich
• OLG Hamburg on LAION and the TDM opt out
• Bartz v Anthropic and the copyright settlement
• Europe AI regulation hub
• AI Regulation News hub