LG München I, Judgment of 31 July 2026 – 42 O 763/25
The Regional Court of Munich I has handed down a significant copyright judgment concerning the generative AI music platform Suno. In its decision of 31 July 2026, the Court largely upheld claims brought by the German collecting society GEMA and found copyright infringements at several distinct stages of the AI system’s development and operation.
The case concerned six well-known musical works, including Atemlos durch die Nacht, Rasputin, Big in Japan, Forever Young, Daddy Cool and the refrain of Mambo No. 5. The dispute did not concern the lyrics of the songs, but rather protected musical elements such as melody, harmony, rhythm and other original features of the compositions.
Three distinct stages of copyright relevance
A particularly important feature of the judgment is that the Court distinguished between three separate stages of possible copyright infringement: the acquisition and reproduction of protected works for AI training, their alleged retention or “memorisation” within the trained model, and the generation of outputs reproducing protected elements of those works.
1. AI training in the United States
The training activities at issue took place in the United States. The Court therefore assessed the relevant reproductions under US copyright law, including the fair use doctrine under 17 U.S.C. § 107.
According to the Court, Suno could not rely on fair use in the circumstances of the case. A central factor was that protected elements of the training works were later reproduced in outputs generated by the system. The Court distinguished the case from US decisions in which training materials were not reproduced in substantially similar form in the outputs.
This aspect of the judgment is especially noteworthy because a German court applied US copyright law to conduct occurring in the United States while adjudicating the wider dispute in Germany.
2. Reproduction within the AI model
The Court went further and held that the protected works were reproduced within Suno’s models in Germany through their storage in parameters or other data structures.
In the Court’s view, a work need not exist inside a model as a conventional audio file in order to constitute a reproduction under § 16 UrhG. It may be sufficient that the protected work can be made perceptible again through technical means on the basis of the model’s internal structure.
The Court inferred such “memorisation” from the relationship between the undisputed training material, the technical functioning of the system and the substantial similarities found in the generated outputs.
It also held that this form of continued reproduction within the model could not be justified by the text and data mining exception in § 44b UrhG.
3. Copyright-infringing outputs
The Court additionally found that certain outputs generated by Suno reproduced recognisable protected elements of the musical works.
Responsibility for those outputs was attributed primarily to Suno rather than to the individual users. According to the Court, the users had not specified the relevant melodies, harmonies, rhythms or arrangements in their prompts. The selection of training data, model architecture and the risk of memorisation remained within the provider’s sphere of control.
The Court therefore granted GEMA substantial injunctive relief and also ordered the defendant to provide information relevant to damages claims.
A broader approach to AI memorisation
The judgment continues an emerging line of reasoning in the Munich court’s AI copyright jurisprudence. In the earlier GEMA v. OpenAI litigation, the Court had already treated the reproducibility of protected song lyrics through simple prompts as evidence that the material had been retained within the model.
The Suno decision extends this approach to musical compositions.
This is potentially more controversial. Similarity between two musical outputs may be more difficult to assess than near-verbatim reproduction of text. Courts must distinguish protectable original expression from commonplace musical elements, stylistic conventions and other unprotected features.
The judgment therefore raises a fundamental question for future AI copyright litigation: Does a model’s ability to generate substantially similar output necessarily demonstrate that a protected work exists within the model as a legally relevant reproduction?
The Munich Court answered that question affirmatively on the facts before it, relying on the evidence as a whole. Whether appellate courts will adopt the same approach remains to be seen.
Significance of the decision
The judgment should not be read as establishing that the use of copyrighted works for AI training is categorically unlawful. Its significance lies instead in the distinction between merely analysing works during training and retaining them in a form that enables recognisable reproduction later.
The case therefore suggests a three-stage framework for copyright scrutiny of generative AI systems:
- reproduction of works in connection with the acquisition and preparation of training data;
- possible reproduction through retention or memorisation within the model itself; and
- reproduction or communication of protected expression through generated outputs.
For AI developers, the decision highlights the legal importance of measures designed to reduce memorisation and prevent the generation of outputs that reproduce protected material.
For right holders, it suggests that output similarity may become an important evidentiary route for challenging what occurs inside otherwise opaque AI models.
The judgment is not yet final. Nevertheless, LG München I has provided one of the most detailed European judicial analyses to date of the relationship between AI training, model memorisation and copyright infringement.