German court bars Suno from using GEMA repertoire

The court prohibited four acts in respect of six compositions. Reproducing those works for training purposes, carried out in the United States. Reproducing them by memorisation in the model itself. Communicating them to the public by offering that model. And reproducing and communicating them through the output the model generates.

Suno argued that a model’s weights are nothing more than mathematically learnt patterns and generalised features. The court did not follow that. It held that recognisable versions of complete works had been stored in the parameters, returning in confusingly similar form on a simple prompt. The fair use defence was rejected. Each further breach carries a penalty of up to 250,000 euros.

The point that reaches furthest is not about music but about jurisdiction. The court held itself competent over training that physically took place in the United States, on the ground that the model’s output reaches European users. Anyone supplying into Europe cannot hide behind the question of where the compute sits.

What does this mean for a business owner playing music in their premises? In the short term, nothing. The ruling targets the model provider, not whoever plays the music. But it does show where the dividing line in this market will fall, and that line is the provenance of the training data.

That is also what our own position rests on. The model supplier behind RadioMatic states publicly that it trained exclusively on rights-free sources, licensed material and commissioned work. That statement is in our legal memorandum, together with our own continuous checks of the output with audio fingerprinting and our own software. Anyone reading this ruling would do well to put that same question to every provider of AI music.

Bird & Bird on Munich Regional Court, case 42 O 763/25