German court finds Suno liable for copyright infringement over AI-generated music

The US company Suno Inc. operates an AI-based music generator that can create complete songs in response to user prompts. Its models were trained in the United States using a dataset comprising millions of full sound recordings, together with information such as song titles, lyrics and musical genres. The training dataset included six musical works from the repertoire of the German collecting society, Gesellschaft für musikalische Aufführungs- und mechanische Vervielfältigungsrechte (GEMA).

These were Atemlos durch die Nacht, Rasputin, Big in Japan, Forever Young, Daddy Cool and the chorus of Mambo No. 5 (A Little Bit of...). The dispute concerned the musical compositions, not the lyrics or rights in the original sound recordings. Suno obtained the relevant audio from YouTube using stream-ripping techniques. According to the court, this involved circumventing YouTube's Rolling Cipher, a technical measure designed to prevent audio and video files from being downloaded.

On 17 January 2025, GEMA sent Suno a cease-and-desist letter and subsequently brought proceedings before the 42nd Civil Chamber of the Munich Regional Court I (Landgericht München I), which specialises in copyright matters. To support its claims, GEMA tested Suno's v3.5 and v4 models using the original lyrics and title of each song, together with a general description of its musical style. The prompts did not specify the melody, harmony, rhythm or arrangement. GEMA submitted 338 prompts relating to the six works and obtained outputs which, it argued, reproduced recognisable parts of the original compositions.

Suno disputed this conclusion. It argued that its models did not store the training recordings but learned mathematical patterns and relationships from them. The company also maintained that any similarities in the outputs resulted from the prompts used by GEMA, particularly the inclusion of the original lyrics, rather than from copies retained inside the model.

The Court heard the case on 9 March 2026. During the hearing, it listened to the six original works and the disputed AI-generated tracks. It rejected Suno's argument that GEMA had deliberately steered the system towards the originals. In particular, the Court found that repeating the same prompt did not progressively direct the model towards a particular result. It also considered that lyrics and general genre descriptions still left the melody, harmony, rhythm and other musical elements open.

On 31 July 2026, the Court largely upheld GEMA's claims in case 42 O 763/25. It prohibited Suno from reproducing the works for the purpose of training its models in the United States and from reproducing them in Germany through their storage in model parameters or other data structures. It also prohibited certain acts of communication to the public in Germany and the reproduction and communication of the disputed AI-generated adaptations. The Court further upheld claims for information and damages in relation to several of the works. The judgment is not yet final, and Suno stated after the decision that it was considering its options, including an appeal.

A central part of the judgment concerns what the Court described as memorisation. It distinguished three stages in the development and use of an AI model: preparing the training material, analysing that material during training and subsequently using the trained model through prompts and outputs. The Court accepted that AI models do not necessarily store conventional copies of their training files. However, it found that the six compositions were present in Suno's models in a form that allowed protected musical elements to be reproduced through relatively simple prompts.

The Court considered this sufficient to constitute reproduction under §16 of the German Copyright Act and Article 2 of the InfoSoc Directive. In its view, a work does not need to be stored as an identifiable audio file. Reproduction can also occur where protected elements are fixed in model parameters or other data structures and can later be made perceptible with the aid of the system.

The Court then addressed the text and data mining (TDM) exception under Article 4 of the DSM Directive and §44b of the German Copyright Act. It did not exclude AI training from the TDM exception as a matter of principle. On the contrary, it accepted that the exception could cover reproductions needed to collect and prepare material for a training corpus.

However, the Court drew a distinction between those preparatory copies and the reproduction of protected material within the trained model. Once protected elements of a work are incorporated into the model in a form from which they can later be reproduced, the Court considered that the process goes beyond extracting information, patterns or correlations. Those reproductions were therefore not covered by the TDM exception.

There was also a separate problem with the way the training material had been obtained. Article 4 requires lawful access to works used for TDM. Since Suno had circumvented YouTube's Rolling Cipher to download the recordings, the Court found that this requirement had not been met.

The Court examined the generated music separately. It found that protected elements of the six compositions remained recognisable in the outputs and treated the disputed tracks as adaptations of the original works. It attributed the resulting reproductions to Suno rather than to the users because the simple prompts did not determine the protected musical features that appeared in the outputs.

In addition, The Court considered the communication of the works to users. It did not uphold GEMA's claim based specifically on the right of making available under §19a of the German Copyright Act. However, it found that Suno's conduct amounted to a broader act of communication to the public under §15(2), interpreted in accordance with Article 3(1) of the InfoSoc Directive. Suno could not rely on the hosting exemption under Article 6 of the Digital Services Act. The Court considered the disputed outputs to be generated by Suno's service rather than information merely stored on behalf of its users.

The case also required the German court to consider US copyright law because the model training had taken place in the United States. The Court found that the copies made for training fell within the reproduction right under 17 U.S.C. §106(1), before examining whether Suno could rely on the four-factor fair use test in §107.

The Court acknowledged that training a generative AI model can, in general, constitute a transformative use. However, it considered the circumstances of this case different. The disputed outputs reproduced substantial protected elements of the same works used for training, while Suno operated the service commercially. The Court also considered the way the recordings had been obtained. It distinguished the case from the 2025 US decisions in Bartz v Anthropic and Kadrey v Meta, where the training works had not been reproduced to a comparable extent in the outputs. After assessing the four factors, the Court concluded that Suno's use of the six works was not protected by fair use.