Suno: The Music AI Found Guilty of Copyright Infringement

Suno: The Music AI Found Guilty of Copyright Infringement

A court in Munich has ruled that the AI music generator Suno infringes copyright by training on well-known musical works, following a complaint filed by the German music rights organization GEMA.

The court found that Suno's AI models had indeed memorized songs and, when prompted, reproduced original elements from the source material, which constitutes a copyright violation. The responsibility for the infringing outputs was assigned directly to Suno rather than its users, as the company chose which training data to use and how to construct its models.

A Munich court determined that the AI music generator Suno violated copyright both through its training process and its outputs. The court also applied U.S. law and rejected Suno's fair use defense. GEMA, the German music rights organization, had sued for an injunction, disclosures, and damages, winning the majority of its claims. The case centered on six well-known songs, including "Atemlos durch die Nacht" by Kristina Bach and "Rasputin" by Frank Farian, Fred Jay, and George Reyam. Only the musical compositions were in question, not the lyrics.

Memorization is Real

The court determined that the six pieces were reproducible in Suno's models 3.5 and 4. In AI research, this phenomenon is known as memorization: during training, a model can not only learn general patterns but also store specific content from its training data, which can then be extracted as output.

Suno argued that its model does not store songs, but only "mathematically learned patterns and generalized features." Any similarity in the outputs, the company claimed, was the result of user prompts and statistical correlations.

For its test, GEMA entered the original lyrics of each song, the musical style, and the title into Suno's generator. No specifications regarding melody, harmony, rhythm, or arrangement were made. Nevertheless, Suno produced results in which the court recognized the original elements of the source pieces. Given the complexity and length of the songs, the court ruled out coincidence.

Suno Holds Responsibility, Not Its Users

The court explicitly held Suno responsible for the infringing outputs, not the users who typed the prompts. Suno had argued that the deliberate input from users broke the causal chain between the model and its output.

The court disagreed. The prompts were "simple and open," and Suno operates the models, selects the songs as training data, and is responsible for the architecture and memorization. This means that the models "substantially determine" the outputs. Simply offering the generator for musical creation already constitutes a legal violation, the court stated. If this position is upheld on appeal, it could also affect other AI music services.

Suno also invoked the German exception for text and data mining, which allows for automated content analysis under certain conditions. The court ruled that this exception does not cover the memorization it observed.

U.S. Fair Use Does Not Protect Suno Either

Under a special rule for management companies, the court also asserted its jurisdiction over claims based on training activities that took place in the United States. It applied U.S. law to these acts and concluded that fair use does not protect Suno either.

The court drew a line between this case and two U.S. proceedings, Bartz and Kadrey. In these two cases, U.S. courts had treated AI training as transformative use and therefore fair use. The key difference, according to the Munich court, is that in those proceedings, the training data was "not, or not substantially, accessible to users in the outputs." With Suno, simple inputs produced outputs "substantially similar" to the originals. All factors of the fair use test established by the U.S. Supreme Court in its Warhol decision weighed against Suno, the court found. The judgment is not yet final.

Key Questions Remain Unanswered

The distinction from the U.S. cases seems broader than it actually is. The Munich court does not say that music models memorize works while text models do not. It simply notes that reproduction was concretely proven with Suno, whereas this was not the case in the other proceedings.

However, research has confirmed this effect with books, suggesting that memorization is not unique to music. The ease with which a work can be extracted from an AI model likely depends on how frequently it appeared in the training data and the specificity of the prompt. Popular works like "Atemlos durch die Nacht" or "Rasputin" are widely available on the Internet, making their memorization more likely. Researchers have also been able to extract well-known books like Harry Potter from language models with high similarity and almost unchanged.

The Munich court's description of GEMA's prompts as "simple and open" seems particularly fragile. GEMA entered the complete lyrics, the musical style, and the title. Although it did not specify musical elements, the identity of the desired song was quite well defined. A typical user creating original music with an AI generator would almost certainly not formulate their prompt in this way.

In the copyright battle between the New York Times and OpenAI, this is a central point of contention: do targeted prompts designed to reproduce protected content reflect normal use of an AI system, or do they represent a special case? In other words, is it a bug when the system produces originals, or a feature?

Suno Allegedly Circumvented YouTube's Protections

No one uses a music generator to recreate a song that already exists, so all the attention on the similarity of outputs could ultimately be a dead end. The more pertinent question is whether Suno was allowed to use copyrighted material for training without consent in the first place.

The court's notes contain an interesting detail on this matter. According to the court's press release, Suno used "stream-ripping techniques" to extract music from YouTube, circumventing YouTube's "Rolling Cipher," a technical protection designed to prevent the downloading of audio and video content.

This raises a question that goes beyond this particular case: does circumventing technical protections already constitute piracy, regardless of what happens next with the data? U.S. courts have so far taken a clear stance on this: fair use may apply to AI training, but not when it is based on pirated material.