AI Music Generator Suno Infringes Copyright Through Memorization

Virtually every organization uses AI nowadays, but what happens when an AI model not only analyzes existing works of others, but also retains and reproduces them? On July 31, 2026, the court in Munich issued a ruling on this matter (Landgericht München July 31, 2026, case number 42 O 763/25 (GEMA/Suno)). The judge ruled that Suno, the provider of an AI music generator, infringes on the copyrights of the rights holders affiliated with the German collective management organization GEMA. The ruling concerns the copyright infringement by Suno. To demonstrate this copyright infringement, six well-known songs were used: Atemlos durch die Nacht, Rasputin, Big in Japan, Forever Young, Daddy Cool, and the chorus of Mambo No. 5.

This is not the first time a German court has ruled in a lawsuit concerning AI and memorization. Last year, GEMA already won a similar case against OpenAI regarding the reproduction of song lyrics (Landgericht München November 11, 2025, case number 42 O 14139/24 (GEMA/OpenAI)). In the lawsuit against Suno, two phases can be distinguished:

• Phase 1 (training the AI model): Analyzing data and detecting statistical patterns for training;

• Phase 2 (memorization of data): Permanently storing or 'memorizing' the protected works in the parameters of the AI model.

This ruling by a German lower court has no precedential effect in the Netherlands, but it does show an interesting development. In this blog, we examine the key considerations of this ruling and what the potential implications are for your organization.

How does Suno work?

Through Suno's AI model, users can generate music based on a text prompt. Users can provide a description of how the song should sound or what the lyrics should be about. The AI model generates a complete song based on the prompt, including vocals, instruments, and a fitting musical style.

To be able to generate music, the AI model analyzed existing music during the training phase. Through this training data, the AI model 'learns' how music structures, rhythms, and melodies work, so it can create new songs (phase 1).

GEMA takes legal action

GEMA took legal action after Suno had not responded to the request to enter into a license agreement for the use of the songs for training the AI model.

GEMA argued that Suno used the songs without permission to train the AI model, and that the songs were unlawfully downloaded from YouTube via stream-ripping (circumventing technical protection measures). Additionally, GEMA claimed that there was unauthorized reproduction of works because the AI model had memorized the melodies, rhythms, and harmonies of the original songs.

Memorization constitutes infringement

Suno defended itself with the argument that the AI model does not store audio files, but only registers abstract statistical patterns. However, the court did not accept this argument. GEMA demonstrated with targeted prompts (the title, original lyrics, and a desired music style of the six well-known songs) that Suno's generated output reproduced the melodies, rhythms, and harmonies of the original songs so recognizably that it could not be coincidental. The court ruled that memorization had occurred.

Memorization occurs when information about specific training data is embedded in the model parameters in such a way that the AI model is capable of reproducing the training data wholly or partially in the output. Whether memorized content is present is determined by comparing the output with the training data: if the characteristic features of the original work are found in the output, this indicates that memorization has taken place. The court ruled in the case between Suno and GEMA that the AI model had effectively stored the copyrighted songs in its parameters. The songs are reproduced almost entirely in the output through simple prompts. This means there is unauthorized reproduction.

It is not necessary for a work or song to exist as a directly accessible file in the AI model. It is sufficient that the work is present in the parameters of the AI model and is made indirectly perceivable and accessible to people through the AI generator. The court ruled in the case between Suno and GEMA that the criterion of accessibility was met because the AI model, in which the original works are stored, was made available to Suno users.

Text and data mining exception

The court also addressed Suno's claim that the AI model falls under the text- and data mining-exception (TDM-exception). The TDM-exception stems from Article 4 of the DSM Directive (Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market) and is established in the Netherlands in Article 15o of the Dutch Copyright Act. This exception permits TDM for research purposes under strict conditions. Whether training of generative AI models falls under this exception is not yet a settled matter. This question remains open.

Furthermore, the court states that retaining training data within the AI model does not fall under the exception. The court rules that Suno's AI model generally falls within the scope of the TDM-exception, insofar as it is used for analyzing and detecting statistical patterns (phase 1 of the training process). However, a protected work that is permanently stored or memorized in the parameters of the AI model does not fall under the TDM-exception (phase 2 of the training process). This retention no longer serves the purpose of further data analysis, but leads to an infringement of the creator's exploitation rights.

YouTube uses 'Rolling Cipher' technology. This technology works with two separate URLs for each audio and video file. The first URL is visible to users and leads to the webpage where the files are played. The second URL is for file storage and is not visible to users. This second URL uses rotating encryption codes that ensure songs can only be listened to and watched, but not downloaded. This security measure was classified by the court as an effective technical measure against unauthorized downloading of songs. The court emphasized that a measure does not need to be 100% foolproof to be considered 'effective.' If that were the case, circumvention would be impossible and legal proceedings against circumvention would be unnecessary.

Suno circumvented YouTube's technical protection mechanisms via stream-ripping. The court therefore ruled that Suno gained unauthorized access to the songs. As a result, Suno cannot invoke the TDM-exception.

What does this mean for you?

For providers of AI models, this means it is important to technically manage memorization risks. Consider techniques such as 'alignment' or 'guardrails' for the AI model and periodic checks comparing output with known training data.

For users, contractual guarantees and indemnifications from AI providers are crucial. However, guarantees and indemnifications do not always provide adequate protection against infringements of third-party copyrights.

For copyright holders, this means a stronger licensing and negotiating position. It should be noted, however, that no ruling has yet been issued on whether AI training with protected works is subject to licensing requirements. Nevertheless, the ruling strengthens their position.

What does this ruling mean for the Netherlands?

Rulings by foreign courts have no direct precedential effect in the Netherlands. Nevertheless, this German ruling may also be relevant for the Netherlands. Copyright law is partly harmonized within the European Union. Certain concepts, such as the 'reproduction right' and the scope of the TDM-exception, are interpreted the same way throughout the European Union. If this lawsuit, or a similar case, leads to preliminary questions to the Court of Justice of the European Union (CJEU), then the answer binds all national courts in the interpretation of EU law.

AI and intellectual property can raise questions. Want to know what this means for your organization? Contact us. We are happy to help!