Three months after we wrote about GEMA’s lawsuit against AI music generator Suno, the court has now delivered its ruling. The German copyright organization has largely won the case, marking an important development in the debate around AI and music rights.

The court found that Suno had used protected songs to train its AI model and that parts of those works could be reproduced in generated output. The case involved several well-known compositions, including Atemlos durch die Nacht, Rasputin, Daddy Cool and Mambo No. 5.

A key part of the ruling is the distinction between an AI model learning general patterns from music and effectively remembering specific works. According to the court, Suno’s model had not only analyzed the songs, but stored them in a way that allowed recognizable parts to appear again in generated music.

That distinction proved important for Suno’s reliance on the text-and-data-mining exception. While such exceptions can in principle allow copyrighted material to be analyzed, the court ruled that they do not cover situations where protected works are effectively reproduced by the model.

The court also rejected Suno’s reliance on US fair use. Unlike in several earlier AI cases involving text, the generated output here could closely resemble the original protected works.

As a result, Suno must stop reproducing and offering the six works involved in the case. The company must also provide information about the extent of the use and revenues generated, and has been held liable for damages. The exact amount still needs to be determined, and Suno may appeal the ruling.

For creators and collective management organizations, the decision strengthens the position that AI companies cannot simply rely on copyrighted music without considering licensing and compensation.

When we first wrote about this case in May, we asked whether AI platforms could continue building on existing music without creators’ consent, or whether they would eventually have to start paying for it.

This ruling does not answer every question surrounding AI training and copyright, but it does move the debate significantly closer to a licensing reality.