Munich court: Suno's AI memorized hits like 'Atemlos,' fair use won't save it

A Munich court just dropped a bomb on AI music, ruling that Suno’s models don’t just learn from songs—they memorize them. The case, brought by German rights organization GEMA, centered on six tracks including the Schlager smash ‘Atemlos durch die Nacht’ and Boney M.’s ‘Rasputin.’ When GEMA fed just the lyrics, style, and title into Suno, the output was a dead ringer for the originals. The court called that infringement, full stop.

Suno tried the usual defenses and got nowhere. Their argument that the model only stores ‘mathematically learned patterns’ collapsed against evidence that specific songs are reproducibly contained in versions 3.5 and 4. The court dismissed the idea that users are to blame for prompting the infringement, placing responsibility squarely on Suno for choosing the training data and building a model that ‘substantively determined’ the infringing outputs. Simply offering the generator, the ruling states, is already a violation.

What makes this ruling travel is its treatment of US law. The court claimed jurisdiction over training acts in America and rejected Suno’s fair use defense, explicitly distinguishing this case from previous wins by AI companies in the Bartz and Kadrey proceedings. The difference? Those cases involved claims of transformative use where the training data wasn’t spat back out to users. Here, the reproduction was proven. It’s a crucial distinction, but a fragile one. The court isn’t arguing text models don’t memorize—research on books like Harry Potter shows they absolutely do—it’s just that nobody proved it in those specific cases.

For now, the ruling isn’t final, and Suno will likely appeal. But the logic is sticky. If a model can be shown to regurgitate training data, the fair use shield gets very thin, regardless of what kind of content it generates. This puts every AI company on notice: what your model outputs can now define your legal liability for what you trained it on.