Munich held Suno liable for what its model produced — and reached the US training that built it

On 31 July 2026 the 42nd Civil Chamber of the Landgericht München I, presiding judge Cornelia Kallert, ruled that the AI music generator Suno infringed copyright in six compositions — “Atemlos durch die Nacht”, “Rasputin”, “Daddy Cool”, “Big in Japan”, “Forever Young” and the refrain of “Mambo No. 5” — and largely granted GEMA’s claims for an injunction, disclosure and a declaration of liability in damages (case 42 O 763/25). Claims over lyrics were excluded from the dispute. The court found the training material had been taken from YouTube using stream-ripping techniques that circumvented its protection measures; that the works were memorised in the model’s parameters, which the text-and-data-mining exception in § 44b UrhG does not cover; and that the generated tracks were themselves infringing reproductions and communications to the public under §§ 15, 16 UrhG, because the original elements of the compositions remained recognisable in them. Memorisation was established by comparing the works against the outputs.

The injunction reaches the training reproductions carried out in the United States as well as the storage of the works inside the model parameters and the provision of the model in Germany, backed by a contempt fine of up to €250,000 per breach. The damages figure is not set — that is what the disclosure order is for. GEMA was represented by Raue, Suno by Latham & Watkins. The judgment is first instance and not final; Suno says it is weighing its options, including an appeal. It follows the same chamber’s ruling against OpenAI over song lyrics (42 O 14139/24, 11 November 2025), which is itself on appeal in Munich — so the doctrine both judgments rest on is still moving.

Why it matters

The holding brand owners should read twice is the one about who answers for the output: the provider of the AI system, not the user who typed the prompt. Every enforcement conversation about AI-generated material — a clone site whose copy was spun from yours, a fake review farm, a synthetic endorsement video — starts with the question of whose door you knock on. A European court has now put the platform behind the door, on a copyright theory, and refused to let the TDM exception carry the model past it.

Counsel’s note

Take the evidentiary method first, because it is the transferable part. The court accepted that a work sits inside a model by comparing the model’s outputs against the work. No discovery into the training set, no insider, no leaked data card — systematic prompting, captured properly, was enough to shift the argument. Any brand that suspects its protected material is inside somebody’s model can build that record itself, and should build it the way it would build any other evidence file: timestamped, reproducible, hashed, collected before the notice goes out rather than after the model gets quietly retrained.

Now the cold part. The reason Munich could reach conduct that happened in the United States is that the claimant was a collecting society, and the route ran through a provision of the German collecting-societies regime that is not available to an individual rights holder. A single brand suing over its own catalogue does not inherit that reach. What it inherits is the German-territory half — model storage, provision of the service, the outputs delivered to German users — which is real leverage, but a narrower target than the headline suggests. Anyone reading this ruling as “offshore training is now actionable by anybody” is going to build a claim they cannot land.

The stream-ripping finding is the quiet one, and for our verticals it is the one that travels furthest. The court treated the circumvention of YouTube’s protection measures as part of the wrong, not background colour. If a scraper strips your geoblock, your paywall, your bot protection or your account gate to harvest your material — for a dataset, for a clone, for an affiliate feed — that is a separate and often easier violation to plead than the downstream copying, in Germany and under the equivalent anti-circumvention rules elsewhere. Log the circumvention at the moment you detect it; that log ages badly if you go looking for it a year later.

Finally, the limit. This was decided on copyright, over six compositions, on the German territorial hooks, at first instance. Most of what actually damages an iGaming, broker or crypto brand through an AI surface is not copyright at all — it is trademark use, impersonation and false attribution, where the doctrine is different and thinner. The ruling gives you a much better answer to “is the platform the right addressee”, and it does not give you a cause of action you did not already have.

What this means for you

If AI-generated material is being used against your brand, address the operator of the model or the surface, not the anonymous prompter, and open the file with output-to-original comparisons captured before you notify anyone. Where the harm runs through an AI answer rather than a page, our guide on removing your brand from AI overviews covers the mechanics, and what to do when the platform ignores your report covers the step after that.