The Regional Court of Munich I largely granted GEMA's claims against Suno on 31 July, finding that six well-known musical works were reproducibly memorised in Suno's v3.5 and v4 models running on German servers and could be extracted through ordinary prompts. The court assigned responsibility for those outputs to Suno rather than to the users who typed the prompts. This is not a ruling that training AI on music is unlawful; it is a ruling about what happens when the training shows through.
Key facts
- The case is 42 O 763/25, decided by the 42nd Civil Chamber of the Landgericht Munchen I on 31 July 2026.
- The works at issue are "Atemlos durch die Nacht," "Rasputin," "Big in Japan," "Forever Young," the chorus of "Mambo No. 5 A little bit of," and "Daddy Cool." Lyrics were explicitly outside the case.
- The court granted injunction, information, and damages claims; no damages figure is published and the judgment is appealable.
- Primary source: the court's own press summary.
What the court actually held
Three findings do the work, and they are more specific than the headlines.
First, on the United States training copies: the court held these infringed under US law and were not fair use, because substantially similar originals were made available through the outputs. It expressly distinguished this from the Bartz and Kadrey cases, where training data was not substantially accessible in what the models produced. That distinction is the hinge. It means the court is not treating ingestion and output as the same act; it is treating output accessibility as what converts one into the other.
Second, on the German servers: the works were reproducibly memorised in the v3.5 and v4 models, which the court treated as an unlawful reproduction not covered by Germany's text-and-data-mining exception. A model file containing a retrievable work is, on this reading, a copy.
Third, on responsibility: prompts containing a title, lyrics, and a style, but no musical instructions, still produced recognisable original musical elements. The user supplied no melody, so the melody came from the model, and the provider answers for it.
How memorisation works
Neural networks are supposed to generalize, learning statistical patterns rather than storing examples. In practice they also memorise, especially material that appears many times in training data. A song that shows up in thousands of copies across the web is exactly the kind of thing a model can reconstruct rather than approximate.
The useful analogy is a session musician with a very good ear. Asked to play something in the style of a hit, they might improvise around it. Asked with the title and the words in front of them, they may simply play the hit, because they know it. The court's finding is that Suno's models, given the title and the lyrics, played the hit. Our lesson on memorisation and hallucination covers the general phenomenon, and training data deduplication covers the standard mitigation.
That is why extractable memorisation, rather than corpus membership, is the legally load-bearing fact. Proving a work was in a training set is an argument about inputs. Getting the work back out is a demonstration.
The acquisition route
The judgment records that Suno obtained the six works by stream-ripping them from YouTube, bypassing its Rolling Cipher protection. The court treated that as part of the factual background rather than issuing a standalone anti-circumvention holding, so it should not be reported as a separate ruling. It is nonetheless the clearest public description of how material entered the pipeline.
Suno's litigation position, as the court records it, was that the weights captured generalized patterns rather than the works themselves, that outputs were not recognisable, that user prompts broke the causal chain, that US training was fair use, and that German uses fell under text-and-data-mining exceptions. The court rejected those arguments on these facts. No post-judgment statement from Suno appears on any first-party Suno channel, so nothing should be reported about an appeal beyond the fact that one is possible.
Why it matters
For anyone building or buying generative audio, the operative question shifts from "what was in your training data" to "can a normal prompt get a protected work back out of your model." That is testable, it is testable by plaintiffs, and it does not require discovery into the training corpus. GEMA's leadership frames the outcome as establishing that AI providers must license the works they use; that is the winning party's characterization, published on GEMA's own site, not independent analysis.
It also connects to earlier reporting without validating it. Our story on the Suno leak that named podcast RSS feeds among collection sources described an alleged pipeline. This ruling adjudicates none of that. What it establishes is why such a pipeline matters when proven work by work: stream-ripped acquisition plus model memorisation plus recognisable output was enough, here, to lose.
The honest caveat: this is one chamber of one German court, on six songs, with no published damages figure, and it is not final. The reasoning is a strong signal about how European courts may approach the question. It is not yet settled law anywhere.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)