A Munich court has ruled that Suno infringed GEMA's music rights by training on protected songs and producing outputs that carried recognizable parts of them. If you run an AI company on unlicensed catalogs, this is the ruling you didn't want to read.
GEMA, the German collecting society for composers, lyricists and music publishers, won most of its lawsuit against Suno on July 31, 2026, when the Munich Regional Court I sided with it on claims for injunctive relief, information and damages. The court's 42nd Civil Chamber said Suno's training dataset contained six protected works from GEMA's repertoire and that those works were reproducible in Suno models v3.5 and v4.
That is the core of the case. The court didn't treat Suno's system as a machine that only learned vague musical style. It found memorisation. According to the Bavarian justice ministry's press release, the chamber said training data can be contained in AI models and extracted as outputs, and that chance could be ruled out because of the complexity and length of the music involved.
The six works were not obscure test files. They included "Atemlos durch die Nacht" by Kristina Bach, "Rasputin" by Frank Farian, Fred Jay and George Reyam, "Daddy Cool" by Frank Farian, "Big in Japan" and "Forever Young" by Marian Gold, Bernhard Lloyd and Frank Mertens, and the refrain of "Mambo No. 5 A little bit of" by David Lubega and Christian Pletschacher. The lyrics themselves were not the legal issue in the Suno case. This fight was about musical works.
Here is the part every AI founder should sit with for a minute. The court said Suno used stream-ripping techniques to extract and copy the songs from YouTube, bypassing YouTube's Rolling Cipher protection. GEMA's prompts, according to the court, gave Suno the original lyrics, the desired musical style and the title, but no melody, harmony, rhythm or arrangement. The chamber still found that original elements of the works were recognizable in the outputs.
The fair use argument failed #
Suno made the arguments you would expect. It challenged copyright protection, disputed recognizability in the outputs, argued that its model stored mathematical patterns rather than training data, and said the US training was covered by fair use. It also argued that user prompts broke the chain of responsibility because the outputs came from targeted user action.
The court rejected the useful part of that defense. It held that GEMA's claims covered reproduction during the US training process, reproduction in models stored on German servers, and the communication and reproduction of outputs in Germany. On US law, the chamber said fair use did not protect Suno because the original works appeared in outputs after simple, open prompts. It also distinguished the case from Bartz and Kadrey, two US AI copyright cases where the training material was not made available to users in substantially similar outputs.
That difference matters. You can argue all day about whether training alone should be legal. Once the output gets too close to the song, the courtroom changes. You are no longer selling abstraction. You are selling a system that can hand back recognizable pieces of someone else's work.
GEMA now has a pattern #
This was not GEMA's first AI win in Munich. On November 11, 2025, the same court largely upheld GEMA's claims against two OpenAI companies in case 42 O 14139/24 over nine German song lyrics, including "Atemlos" and "Wie schön, dass du geboren bist." The court said the lyrics had been memorized in GPT-4 and GPT-4o and could be reproduced through simple prompts. It ordered injunctive relief, information and damages in principle. An appeal is pending.
Two rulings do not settle all of AI copyright law. They do give rights owners a German template. GEMA has now shown that a collecting society can bring a model memorisation case, point to specific works, compare outputs with originals, and win at the trial court level. The Suno judgment is not final, so the company can still fight it higher up. But the first order is already there.
The business context makes the ruling harder to shrug off. Music Business Worldwide reported that Suno raised more than $400 million in a Series D round on June 3, 2026, at a $5.4 billion post-money valuation, more than double the $2.45 billion valuation it reached after its Series C in November. Investors put serious money behind a company that can generate full songs from text prompts in seconds. A Munich court has now said the legal basis for part of that machine fails in Germany.
GEMA says it represents more than 100,000 members in Germany and more than two million rights holders worldwide. That scale is the point. If licensing becomes the practical price of operating an AI music model in Europe, Suno's cost base changes from compute and growth to permissions as well. Pay for the catalog, or keep it out of the model.
Suno is not finished. The Munich ruling leaves its US lawsuits untouched, and Germany is not done with it yet. But you should not miss what has shifted. The company can still say its models create new music. In Munich, GEMA showed the court six songs and got a different answer.
Also read: OpenAI's unreleased Astra model solved ten open math problems for $2,000 • Huawei Open Sources a 505 Billion Parameter AI Model Built Without Nvidia Chips • Executive Order 14409 Just Set the Line for Dangerous AI Models