On 31 July 2026, the 42nd Civil Chamber of the Regional Court of Munich I ruled largely in favour of the German collecting society GEMA against Suno, in case 42 O 763/25. The chamber specialises in copyright.
The finding at the centre of it is not about training in the abstract. It is about what the court held to be sitting inside the finished model.
The sentence the case turns on
From the court’s own press release, in the chamber’s reasoning:
“Nach Überzeugung der Kammer seien die streitgegenständlichen Musikstücke reproduzierbar in den Modellen Version v3.5 und v4 der Beklagten enthalten.”
In the chamber’s judgement, the songs at issue are reproducibly contained in the defendant’s v3.5 and v4 models. The training run did not consume them. They can be brought back out of the finished model.
The court’s route to that conclusion is worth following, because it is evidential rather than theoretical. It notes that research in information technology has established that training data can remain inside AI models and be extracted as outputs, a phenomenon known as memorisation. It compared the songs in the training data against the reproductions in the outputs. And it held that given the complexity and length of the pieces, coincidence was excluded as an explanation.
The models were stored on servers in Germany, which is how the court reached jurisdiction.
Six songs, named
This is not a case about a category of works. The judgment concerns six named pieces: “Atemlos durch die Nacht” by Kristina Bach, “Rasputin” by Frank Farian, Fred Jay and George Reyam, “Big in Japan” and “Forever Young” by Marian Gold, Bernhard Lloyd and Frank Mertens, the chorus of “Mambo No. 5 (A Little Bit Of)” by David Lubega and Christian Pletschacher, and “Daddy Cool” by Frank Farian.
Six works, six specific comparisons between training input and generated output. That specificity is why the coincidence argument failed.
What was infringed, and what did not save it
Reproduction right, section 16 UrhG. Through memorisation, the court found the reproduction right infringed by the presence of the works in the models themselves.
The text and data mining exception does not cover it. The chamber held that this reproduction inside the models is not covered by the TDM limitation in section 44b UrhG. That provision is the one most often cited as the legal basis for training on copyrighted material in the EU.
The outputs are a second infringement. The court further held that by reproducing the songs in outputs in Germany, the defendants infringed the works again, without authorisation.
A fair use defence was considered and rejected. The chamber examined the factors required by the US Supreme Court’s Warhol decision and found that all of them told against the defendant. A German court working through American fair use factors is unusual, and it happens when the defendant raises them.
The court awarded injunctive relief, information and damages, granting GEMA’s claims for the most part.
What is not established
The judgment is not final. The court’s press release says so in as many words: “Das Urteil ist nicht rechtskräftig.” Suno has said it disagrees and is considering an appeal, and the damages amount is still to be determined.
Provisional enforceability is not documented. A figure circulates in search results claiming the judgment is provisionally enforceable against security of 150,000 euros per claim. The court’s press release says nothing about it. We looked. If you see that number, it did not come from this source.
That distinction matters more than it sounds, because provisional enforceability is exactly the sort of detail that decides whether a company has to change its product now or after an appeal.
Why the memorisation finding travels
Most coverage of AI copyright disputes argues about the training stage: whether ingesting works to learn from them is an infringement or an exception. This ruling steps past that argument.
If the works are reproducibly contained in the shipped model, the model itself is the copy. That reframes the question for anyone deploying a model they did not train: the exposure does not end when the training run does, and “we only used the weights” stops being a defence in the way it was assumed to be.
The second finding compounds it. An infringement in making outputs available attaches to whoever runs the service, not to whoever built the model.
Reading a first-instance ruling correctly
A first-instance judgment is a real event with consequences for the parties, and it is not settled law. Both halves get quoted, and only one of them survives an appeal.
What this establishes today: a specialist copyright chamber, applying an evidential test for memorisation, found the TDM exception inapplicable and a fair use defence unpersuasive, in a case with named works and a case number you can look up.
What it does not establish: that the reading holds on appeal, that other German courts share it, or that it transfers to other kinds of model.
“German court rules AI training illegal” fits a headline, drops the word that changes its meaning, and misses the more interesting finding underneath.
Sources
- Landgericht München I, press release on the judgment GEMA v. SUNO, 31 July 2026, published by the Bavarian State Ministry of Justice: https://www.justiz.bayern.de/gerichte-und-behoerden/landgericht/muenchen-1/presse/2026/16.php (retrieved 27 August 2026). Source of the case number 42 O 763/25, the chamber and ruling date, the six named works and their authors, the quoted reasoning on reproducible containment in models v3.5 and v4, the memorisation finding and the exclusion of coincidence, the German server location, the infringement of section 16 UrhG, the inapplicability of section 44b UrhG, the second infringement through outputs, the rejected Warhol fair use factors, the awards of injunctive relief, information and damages, and the statement that the judgment is not final. The press release makes no statement about provisional enforceability or security amounts; that absence was checked, not assumed. Quotations verified against the page source.
- Gearnews, “GEMA vs. Suno: Munich Court Now Rules AI Music Training Illegal”, 31 July 2026: https://www.gearnews.com/gema-vs-suno-tech/ (retrieved 26 August 2026). Source of Suno’s stated disagreement and that it is considering an appeal, of the damages amount remaining to be determined, and of the chamber’s earlier narrow reading of the TDM exception in its November 2025 OpenAI ruling.
