GEMA wins
also against
Suno.
GEMA wins
also against
Suno.
of
Does artificial intelligence simply learn, or does it store music created by others? And what does it mean when a German court rules on training that took place in the United States?
Six songs and a music generator
As a collecting society, GEMA administers the rights of composers, lyricists, and music publishers. It had filed a lawsuit against Suno , the U.S.-based provider of the AI music generator, seeking an injunction, disclosure of information, and damages. According to GEMA, the lawsuit was filed after a request for licensing went unanswered.
The subject of the dispute was six compositions, namely “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 chorus of “Mambo No. 5 A Little Bit of” by David Lubega and Christian Pletschacher, and “Daddy Cool” by Frank Farian. Infringements caused by the song lyrics were expressly not the subject of the proceedings. The case therefore concerned the music, not the lyrics.
According to the court’s findings, Suno had downloaded these works from YouTube using stream ripping, thereby circumventing a technical protection measure called Rolling Cipher, which is designed to prevent downloads. GEMA employees entered the respective original text, the desired music style, and the title of the work into the music generator to preserve evidence, in some cases multiple times—for example, 176 times for “Atemlos durch die Nacht.” GEMA submitted the audio recordings generated in this way to the court as evidence. In January 2025, it issued a cease and desist letter to Suno, after which a lawsuit was filed.
Learn or Copy?
GEMA argued that the works were fully stored in the model—that is, defined in the model’s parameters in such a way that they could be reproduced at any time. This, GEMA contended, constituted a permanent reproduction and not merely an analysis of training data. Suno countered that the model did not store any works, but merely mapped statistical probabilities for sound patterns. Furthermore, the training took place in the United States and was covered there by the fair use doctrine.
The Judgment
The Munich I Regional Court has, in Judgment of July 31, 2026 – Case No. 42 O 763/25 The court largely granted GEMA’s motions for an injunction, disclosure, and a declaration of liability for damages. Suno must refrain from reproducing the six works without GEMA’s consent, using them as the basis for training, and publicly performing them via the music generator.
For collecting societies, music publishers, and individual authors, the decision to use artificial intelligence is significant for several reasons. It affects not only the training process in the United States but also the question of whether a trained model itself contains a copy.
Why Training Itself Is a Form of Reproduction
The court relies heavily on the phenomenon of memorization, which is well known in AI research. According to this, memorization occurs when a model, during training, not only learns general patterns from the data but also adopts the content of a specific work in such a way that it can later be reproduced in a recognizable manner. According to the court, whether this is the case can be determined by comparing the training data with the generated outputs. Since the inputs used did not contain any musical specifications and chance could be ruled out given the complexity of the generated pieces, the reproduction can only be explained by memorization within the model.
The musical works at issue are reproduced in the models. Storing them constitutes a reproduction for the purposes of copyright law.
In the Chamber’s view, the exception for text and data mining—which permits the automated analysis of copyrighted works under certain conditions—did not help Suno simply because there was no lawful access to the works. Anyone who gains access by circumventing an effective technical protection measure cannot invoke this exception.
Why is a Munich court allowed to rule on events in the U.S.?
The Chamber also held that it had international jurisdiction over the acts of infringement committed within the United States. This is based on a provision of the Collecting Societies Act, which establishes a special venue based on the factual connection for disputes involving a collecting society. In the Chamber’s view, this provision covers not only local but also international jurisdiction.
With regard to the acts committed in the United States, the Chamber applied U.S. law in accordance with the country-of-origin principle but nevertheless did not consider the reproductions for training purposes to be covered by the fair use doctrine.
Ultimately, the court rejected the fair use defense. It held that the works produced by Suno did not serve a new, transformative purpose, but rather fulfilled the same purpose as the originals, namely listening to the music. Furthermore, the use was commercial in nature, in part because Suno had circumvented the protection mechanism to gain free access that a paying user would otherwise have had to pay for. Finally, the court found that there was a negative impact on the market for the original works, as the generated music tracks competed with the originals as a substitute offering.
For rights holders, this opens the door to bringing a case before a German court even if the decisive technical act takes place abroad. It remains to be seen whether this reasoning will hold up on appeal.
Where GEMA Failed to Gain a Foothold
GEMA was unable to fully enforce its claims under the law of making works publicly available, which typically applies to on-demand services where users can access a work at any time. The court considered this requirement unproven because, in some cases, up to 176 identical entries were required to access the works at issue. This did not constitute access at any time of the user’s choosing. However, GEMA had also based its claims, in the alternative, on the more general law of public performance, on which the claim for injunctive relief could ultimately be based. Accordingly, the remainder of the complaint was dismissed.
Who is responsible for the output—Suno or the user?
The Munich judges assigned responsibility to Suno and not to the users. A key factor was that the outputs were generated by simple, open-ended prompts that specified only the song lyrics and musical style. Suno operates the models, selected the musical pieces as training data, and is responsible for the architecture and the learning process. Thus, the models determined the content of the output. The original elements of the musical pieces are recognizable in the output.
In addition, the Chamber considers that the mere offering of the model and the application for music creation constitutes an infringement of an unspecified law of public performance. This, too, is a development that favors rights holders, as it does not require proof of individual user actions.
Classification
Back in November 2025, the same chamber of the Munich I Regional Court had ruled in favor of GEMA against OpenAI because ChatGPT reproduced the lyrics of well-known songs based on simple inputs. OpenAI has appealed this ruling, and the case is pending before the Munich Higher Regional Court. The Suno ruling now extends the line of reasoning regarding memorization from song lyrics to complete musical compositions for the first time and is therefore likely to be significant for the entire industry of AI-powered music generators.
For providers of AI systems, the ruling poses a significant risk if training data is obtained without a clear chain of rights, particularly when technical protection measures are circumvented. For collecting societies and other internationally active rights holders, the decision demonstrates that cross-border infringements may, under certain circumstances, be consolidated before a single German court.
What Rights Holders Can Take Away From This
For legal enforcement, it is crucial to prove that one’s own work appears in a system’s output—and that this occurs in response to inputs that are as simple as possible. The simpler the input, the harder it is to argue that the user produced the result on their own. Anyone documenting such output should save the input, the time, the model version, and the result together.
It is also worth taking a closer look at how the training data was obtained. If a technical protection measure was circumvented, this adds a separate basis for liability in addition to the copyright assessment.
Ultimately, the claims remain work-specific. The ruling concerns six specific compositions.
Conclusion
The outcome was not surprising following the ruling against OpenAI. What is noteworthy is not so much that the chamber found an infringement, but rather where it located the infringement—namely, in the finished model as a permanent medium for protected content. This liability cannot be shifted to users through terms of use.
If the decision stands, it will be favorable for rights holders. International jurisdiction in this case is established under the Collecting Societies Act. This does not serve as a basis for individual authors themselves.
The issue of memorization will likely continue to influence decisions regarding copyright infringements by AI for some time to come.
We are happy to
advise you about
AI!
