German court rules AI music generator Suno broke copyright law - what does that mean?

On July 31, a court in Munich ruled that the AI song generator, Suno (a US-based company), broke copyright law by training on protected music without permission. It's one of Europe's first binding decisions saying AI companies must license the songs they learn from, not just take them. The ruling could give songwriters and labels real leverage in licensing talks. 

But wait, isn’t Suno a US-based company? How can a German plaintiff sue it for something done in the US?

Yes, Suno is a Delaware-incorporated company that is headquartered in Cambridge, Massachusetts. However, the long arm of the German court used a special provision called VGG (the Verwertungsgesellschaftengesetz) that gave GEMA, a German organization that licenses music rights on behalf of composers, lyricists, and publishers, the right to treat the German court as a proper place to file its lawsuit, even against a foreign company for conduct that partly happened abroad. 

So how did Suno break the copyright law?

Four ways, actually. 

First, the court found Suno obtained the songs by "stream-ripping" them from YouTube, circumventing a technical protection measure called a "rolling cipher" that is meant to prevent downloading. So the very acquisition of the training material was treated as unlawful. 

Second, the court found Suno’s model “memorized” or stored the songs in dispute, not just abstract statistical patterns. The court inferred this because the similarities between the originals and Suno's outputs were too extensive and specific to be coincidental.

Third, the court held Suno, not the AI users, responsible for the infringement, reasoning that users only entered basic prompts while Suno designed, trained, and operated the system that produced the infringing output. 

And lastly, simply offering the model and the music-generation application to the public was itself found to infringe the right of public performance under German copyright law.

But isn’t training an AI model considered fair use?

This is where the case differs from some recent US rulings. In Bartz v. Anthropic and Kadrey v. Meta (both June 2025), courts sided with the AI companies, finding that training on authors' works was "transformative" and a fair use, in part because the models were not reproducing the original works in their outputs. But in Suno’s case, the model was outputting works that were substantially similar to the original songs, so fair use did not apply.

So it does matter how similar the AI outputs are to the original work.

What are key takeaways from this ruling?

1. The fair-use argument that AI companies rely on is not a given. The same training behavior can be legal in one major market and illegal in another, and any AI company operating globally now has to reckon with the stricter European line.

   
   

2. AI training is not automatically transformative. The argument that training turns original works into unrecognizable abstract data and math fails if it can be shown that the models retain the original works and can produce something substantially similar.

   
   

3. European rights holders have reach. This case serves as a template that other European rights organizations can copy to bring US companies to court.