Suno Lost in Munich. What the Ruling Actually Says — and What It Changes for You.

On 31 July 2026 the Regional Court of Munich I ruled against Suno, the AI music generator, in a case brought by GEMA, the German collecting society that represents composers, lyricists and music publishers. The case — 42 O 763/25, heard by the court’s 42nd Civil Chamber and filed on 21 January 2025 — turned on six works: “Forever Young” and “Big in Japan” by Alphaville, “Mambo No. 5” by Lou Bega, “Atemlos durch die Nacht” by Helene Fischer, and “Daddy Cool” and “Rasputin” by Boney M. The court ordered Suno to stop reproducing those works, to stop using them to train its model, to disclose the revenue connected to the infringement, and to pay damages in an amount still to be determined.

The finding underneath the headline is narrower and more interesting than “AI is illegal.” The court did not rule on learning in the abstract. It ruled on memorisation — the songs, it held, were reproducibly contained in Suno’s models, specifically versions v3.5 and v4. Presiding judge Elke Schwager put the evidentiary point plainly: given the complexity and length of the musical pieces, randomness as the cause of their playback can be ruled out. In other words, the court was not persuaded that the system had merely learned general patterns about how pop songs work. It concluded the works themselves were retained, and could be pulled back out.

Why the two standard defences did not hold

The first defence was the European text and data mining exception, the provision AI developers routinely point to when asked what entitled them to the training set. The court rejected it, holding that the exception covers the analytical phase of training — not the retention of works inside the model in reproducible form. This is the second time the same Munich chamber has said so. On 11 November 2025, in case 42 O 14139/24, it found against OpenAI on materially the same logic: memorised song lyrics amounted to reproduction, and the TDM exception reaches only preparatory acts that do not exploit the rightsholder’s interests. That earlier judgment is under appeal. Two rulings, one chamber, nine months apart, pointing the same direction.

The second defence was geography. Suno’s training happened in the United States, which is normally where this argument ends — different country, different law, not a German court’s problem. The Munich court did not stop there. It applied US law to the US training copies and ran a full fair use analysis under 17 USC § 107, concluding the copies were not covered, because simple prompts produced outputs substantially similar to the originals. German law then governed the reproductions stored on German servers and made available in Germany. That combination — train abroad, reproduce and serve locally — is the part of this ruling with the longest reach, and it is why lawyers reading it are telling clients to document how their training data was lawfully obtained and to honour machine-readable rights reservations rather than treat memorisation as a theoretical risk.

What it means if you are the one putting music in an ad

First, the honest caveat: this judgment is not final and not yet enforceable. Suno says the verdict rests on a fundamental mischaracterisation of how its technology works, and is evaluating all available options, including an appeal. So it is not settled law, and anyone telling you it is has got ahead of the record. What it is, is a repricing. GEMA’s chief executive, Dr Tobias Holzmüller, framed the outcome as AI providers having to pay for licences instead of helping themselves to members’ works — and a court in a major European market has now said that framing is enforceable enough to carry an injunction, a disclosure order and damages.

Second, the licensed route already exists, and it is being built in public. Warner Music Group settled with Suno on 25 November 2025, licensed its catalogue, sold Songkick to Suno as part of the deal, and secured controls for its artists over whether their names, images, likenesses, voices and compositions are used in AI-generated music; Suno said licensed models would replace its current ones during 2026. But a deal with one rightsholder covers that rightsholder. GEMA’s repertoire was not in it — which is precisely why this case still had to be tried. “Licensed” is not a property a tool has. It is a property of a specific catalogue, in a specific territory, at a specific version of the model.

So the useful question changed shape. It is no longer “is AI music legal.” It is: which repertoire is this specific output licensed against, in which countries will it run, on which model version was it made — and if the answer turns out to be wrong, whose name is on the campaign? That last one rarely belongs to the tool. Worth noting too that the same week carries a separate deadline: the EU AI Act’s transparency obligations for synthetic audio apply from 2 August 2026, which is a labelling duty rather than a licensing one, but lands on the same desk.

None of this makes generators useless. It makes them a place to find an idea rather than a place to source a master. When a piece has to clear legal in more than one country, the unglamorous answer keeps winning: music performed by people, recorded in a room, delivered with the rights written down and the paperwork attached. That is what Onyx Live Strings is — a real string section, real players, and a master you can actually point to a contract for. Tell us what the piece has to do and where it has to run, and we will tell you what it takes to own it outright.