A Munich court has handed GEMA a significant win against Suno, finding that the AI music generator violated copyrights in a case focused on well-known musical works. The ruling is not yet final, but its reasoning goes directly to one of the central questions facing generative AI: when a model can reproduce protected material, who is legally responsible?
The dispute centered on six songs
GEMA, Germany's music rights organization, sued Suno for injunction, disclosure, and damages. The organization prevailed on most counts.
The case concerned six well-known songs, including "Atemlos durch die Nacht" by Kristina Bach and "Rasputin" by Frank Farian, Fred Jay, and George Reyam. The dispute was about the musical compositions, not the lyrics.
The court found that Suno violated copyrights through two connected parts of its system: the training process and the outputs produced by the service. That matters because the ruling did not treat the generated songs as isolated user activity. It treated the model, the training choices, and the resulting output as part of the same legal chain.
Why memorization became the key issue
The court determined that all six tracks are reproducibly contained in Suno's version 3.5 and 4 models. In AI research, this is known as memorization: a model may absorb specific content from training data, not only general patterns, and that content can later appear in outputs.
Suno argued that its models do not store songs. According to the company, the system keeps only "mathematically learned patterns and generalized features." It also argued that similarities in generated music came from user prompts and statistical correlations.
The court was not persuaded. GEMA tested the system by entering each song's original lyrics, musical style, and title. It did not specify melody, harmony, rhythm, or arrangement. Even so, Suno generated results in which the court recognized original elements of the source tracks.
Because the songs were complex and lengthy, the court ruled out coincidence. That finding turned memorization from a technical concept into a legal problem: if protected musical elements can be drawn out of the model, the model is not merely learning broad musical style in the court's view.
Suno was held responsible for the outputs
A central part of the ruling is that the court placed responsibility on Suno, not on the people entering prompts. Suno argued that user input broke the causal chain between the model and the final output.
The court disagreed. It described the prompts as "simple and open-ended" and emphasized that Suno operates the models, selected the songs as training data, and is responsible for the architecture and the memorization. In the court's wording, the models "substantively determined" the outputs.
That reasoning gives the case broader importance. If a provider controls the model, the training data, and the way the system generates music, the court's view is that the provider cannot shift responsibility to users simply because a prompt was involved.
The court also said that offering the generator for music creation already constitutes a legal violation. If that position holds up on appeal, it could affect other AI music services as well.
Fair use and data mining defenses failed
Suno also pointed to Germany's text and data mining exception, which allows automated analysis of content under certain conditions. The court ruled that the exception does not cover the memorization it found in this case.
The court also addressed activity that took place in the United States. Under a special rule for collecting societies, it claimed jurisdiction over claims based on US training activity, applied US law, and concluded that fair use did not protect Suno.
The court distinguished the case from Bartz and Kadrey, two US proceedings where American courts had treated AI training as transformative use and fair use. The Munich court said the difference was that, in those cases, the training data was "not, or not substantially, made accessible to users in the outputs." With Suno, simple inputs led to outputs that were "substantially similar" to the originals.
The court found that all factors of the fair use test laid out by the US Supreme Court in its Warhol decision weighed against Suno.
What remains unsettled
The ruling does not say that music models always memorize works while text models do not. It says reproduction was concretely proven with Suno, while that was not the case in the other proceedings the court discussed.
The source article also notes that research has confirmed a similar effect with books. How easily a work can be extracted from a model likely depends on how often it appeared in training data and how targeted the prompt is. Popular works such as "Atemlos durch die Nacht" and "Rasputin" are widely available online, making memorization more likely.
There is also a tension in how the court described GEMA's prompts. Although GEMA did not specify musical elements, it entered the complete lyrics, the musical style, and the title. That made the identity of the intended song clear. The broader question is whether prompts designed to reproduce protected content reflect normal use of an AI system or a special test case.
The training source may prove just as important as output similarity. According to the court's press release, Suno used "stream-ripping techniques" to extract music from YouTube and bypassed YouTube's "Rolling Cipher," a technical safeguard intended to prevent downloading audio and video content.
That detail raises a separate issue: whether bypassing technical protections is already a problem, regardless of what the model later produces. The source article notes that US courts have so far taken the position that fair use may apply to AI training, but not when the training is based on pirated material.