AI music generators: Suno judgment strengthens rights holders' position

The judgment of Munich I Regional Court dated 31 July 2026 (case no. 42 O 763/25) against the AI music generator Suno is one of the most significant copyright decisions to date concerning generative AI. The court not only turns its attention to the output generated but also assesses how the models are trained, how protected works are stored and how far the provider's liability extends. For companies that develop or use generative AI, the judgment provides important guidance on licensing, data provenance and compliance.

In its judgment, the court largely upheld GEMA's claims against Suno for injunctive relief, disclosure of information and compensation. The subjects of the dispute were six well-known musical works ("Forever Young", "Atemlos durch die Nacht", "Rasputin", "Big in Japan", "Daddy Cool" and the chorus of "Mambo No. 5 (A Little Bit of...)". The issue was not about any infringement regarding the lyrics, but about the composition of the works themselves. The proceedings concerned the Suno models versions v3.5 and v4. The judgment is not yet final and absolute.

Munich I Regional Court distinguishes between training, memorisation and output

The 42nd Civil Chamber assesses the training of the models in the USA, the memorisation in the models stored in Germany and the generation of the outputs separately in each case. It also treats making the model available as a discrete act of public performance.

According to the court's findings, Suno extracted the six musical works from YouTube by means of stream ripping for training purposes, circumventing the so-called "rolling cipher". The company used the files obtained in this way to train its AI models in the USA. In the Chamber's view, the reproductions created in this way were not covered by the US 'fair use' doctrine.

The trained models v3.5 and v4 were then stored on servers in Germany. Because the models generated songs in response to simple prompts in which original elements of the original compositions were recognisable, the court concluded that the protected works were "memorised" in the models, meaning that they were contained in them in a form that could be reproduced. It regarded this as an independent reproduction within the meaning of section 16 German Copyright Act (UrhG). In the Chamber's view, the songs generated also infringed copyright: Although they were not exact copies of the original works, they did contain recognisable, protected elements.

Targeted prompts could shift control of the situation

The Chamber examined whether the generated songs incorporated protected elements by comparing the output with the original works. It ultimately regarded the original elements of the musical works as recognisable to an average music listener. The wording of the prompts was not the determining factor in this assessment, though it was in the subsequent question of to whom the outputs should be attributed. In each case, the claimant had only entered the original song lyrics, the desired musical style and the title of the work; the prompts did not contain any specifications regarding melody, harmony, rhythm or arrangement.

The Chamber attributed the copies resulting from the outputs to Suno on account of the company having selected the training data, having determined the model architecture and being responsible for the memorisation. Suno did not lose control over the situation simply by submitting the prompts to users. In the court's view, the situation might be different if a user exercises targeted control of the output by entering repeatedly modified, evaluative or suggestive prompts. The prompts at the centre of the dispute, by contrast, were simple and open-ended; the specific musical content would therefore have been determined not by the users, but by the models operated by Suno.

Munich I Regional Court also takes AI training in the USA into account

The international scope of the decision is also noteworthy. In any case, AI providers cannot avoid scrutiny from German courts in relation to actions brought by collecting societies simply by having the training take place in the USA. In the Chamber's view, section 131 (2) German Collecting Societies Act (VGG) permits claims relating to acts of use within Germany and acts of training in the US against the same infringer to be consolidated before one German court. However, for the training in the USA, the Chamber applied US copyright law in accordance with the country-of-origin principle.

For providers, this is also relevant beyond the individual case. The Chamber does not rule out the fair use defence provided for under US copyright law in the abstract for all AI training, only in relation to the specific circumstances of use. It holds that there are four crucial factors to consider: the purpose and nature of the use, the nature of the copyright-protected work, the extent and significance of the part used in proportion to the copyright-protected work as a whole and the impact of the use on the potential market or the value of the copyright-protected work. Anyone who uses copyright-protected works for training purposes cannot, in any event, simply appeal to fair use if the model reproduces the protected works – or at least their original elements – in a substantially similar form.

Memorisation in the model: section 44b German Copyright Act (UrhG) does not apply

The Chamber first clarifies that the text and data mining exception under section 44b German Copyright Act (UrhG) is, in principle, also applicable to the training of generative AI. However, it only permits reproductions that are necessary to perform automated analyses of digital works, such as when creating a training corpus. In the Chamber's view, permanently incorporating protected elements of a work into the model parameters does not serve the purpose of further data analysis and is therefore not covered by section 44b German Copyright Act (UrhG). Irrespective of this, there is no lawful access to the works because Suno downloaded them by circumventing YouTube's rolling cipher.

Communicating to the public even without reliable access

In addition, the court considered the issue of communication to the public. In the GEMA/OpenAI case, the Chamber had affirmed that the work had been made available to the public under section 19a German Copyright Act (UrhG). In this case, it denied that the work had been made available to the public on the grounds that the musical works were not reliably available and therefore could not be accessed at a time of the user's choosing. In the Chamber's view, however, the provision of the model and the outputs in dispute nevertheless infringed the unspecified right to communicate to the public as set out in section 15 (2) German Copyright Act (UrhG). The judgment thus builds on the model deployer's liability for communication to the public, as already recognised in the GEMA/OpenAI judgment, but in this case bases it on a different legal foundation.

Suno judgment increases pressure on licensing and compliance

The Munich Chamber is thus continuing the approach it adopted in the context of generative AI and applying it to AI music generators. For rights holders, the judgment strengthens their ability to take legal action against providers in Germany if models are stored here or made available in Germany. Conversely, providers of generative music systems are facing increasing pressure to provide robust clarification on licensing models, data sources, technical safeguards and internal documentation.

Conclusion: What the Suno judgment means for AI providers

• Munich I Regional Court regards training, memorisation, the provision of models and output as separate levels of copyright assessment.
• Liability may arise even during the training and storage of protected works and is not limited to the output generated.
• Only prompts designed to guide the model in a specific way could shift control to the user; simple, open-ended prompts are not sufficient for this purpose.
• For providers of generative AI, licensing, the origin of training data and technical documentation are becoming increasingly important. 

The judgment is likely to have a lasting impact on the development of compliance and licensing strategies for AI music generators.