German Suno ruling puts AI music licensing under the spotlight

The debate over copyright and generative AI has moved another step closer to the courts determining where the boundaries lie.

On 31 July, the Munich Regional Court I largely upheld claims brought by German collecting society GEMA against Suno, the US-based company whose generative AI platform allows users to create complete songs from text prompts.

The case concerned six works represented by GEMA: ‘Atemlos durch die Nacht’, ‘Rasputin’, ‘Big in Japan’, ‘Forever Young’, the refrain of ‘Mambo No. 5’ and ‘Daddy Cool’. The court upheld claims for injunctive relief, information and damages, although the amount of damages has yet to be determined. The ruling is not yet final.

Suno has said it disagrees with the judgment and is considering its options, including an appeal.

More than a training-data case

The importance of the case lies partly in what the court concluded had happened inside Suno's models.

According to the court, the six works had been included in Suno's training dataset, having been extracted from YouTube using stream-ripping techniques. It found that versions of the works were subsequently ‘memorised’ within Suno's v3.5 and v4 models and could be recovered through generated outputs.

GEMA's test prompts supplied the original lyrics, title and desired musical style, but did not specify melody, harmony, rhythm or arrangement. The court concluded that the musical similarities appearing in the resulting outputs could therefore be attributed to information retained by the model rather than simply to detailed instructions supplied by the user.

That distinction matters.

The judgment should not necessarily be read as establishing that every act of training a generative AI system using copyright material is automatically an infringement. Instead, the Munich court focused heavily on the finding that the works had not merely been analysed to identify patterns, but had been retained sufficiently within the model for recognisable elements to reappear in outputs.

The court found that this form of memorisation constituted reproduction under German copyright law and was not protected by Germany's text and data mining exception.

In its analysis of the judgment, international law firm Bird & Bird notes that the ruling potentially leaves a different position for models that do not memorise protected works, or where similarities arise only after highly directed prompts.

A ruling that reaches across the Atlantic

Potentially still more significant is the international dimension.

Although Suno's training took place in the US, the Munich court held that it could consider GEMA's claims relating to those activities and applied US copyright law to them.

Suno had argued that its training activity was protected by the US doctrine of fair use. The court rejected that defence in relation to the works before it.

In doing so, it distinguished the Suno case from recent US cases involving Anthropic and Meta, where courts had considered AI training in circumstances where the underlying works did not substantially reappear in model outputs. In the Suno case, the Munich court found that substantially similar music could emerge from relatively simple prompts, leading it to conclude that the circumstances were materially different.

The court also considered Suno's commercial purpose, the creative nature of the works, the fact that complete works had been copied and the potential effect of competing AI-generated outputs on the market when assessing fair use.

GEMA CEO Dr Tobias Holzmüller described the judgment as an important step for creators, arguing that AI providers should license the works on which their commercial systems depend. GEMA says the ruling strengthens the ability of European rightsholders to pursue claims even where model training takes place elsewhere.

Why the decision matters beyond Germany

The judgment comes as governments and courts around the world wrestle with essentially the same question: how should copyright law apply when generative AI systems learn from existing creative works?

In the UK, the issue remains unresolved. Following extensive consultation, the UK Government said in March that its previous proposal for a broad copyright exception accompanied by an opt-out was no longer its preferred approach. It is instead gathering further evidence on licensing, transparency, enforcement and the effects of copyright rules on both creators and AI developers.

The German decision therefore arrives at a particularly important moment.

For music publishers, composers and collecting societies, it strengthens the argument that commercial AI development cannot simply exist outside the established system of copyright licensing.

For AI companies, however, its most important lesson may be more specific: memorisation and reproduction could prove legally very different from genuinely transformative machine learning.

That distinction may become increasingly important as courts develop a more sophisticated understanding of what actually happens inside generative models.

The Munich ruling will not settle the global argument over AI and copyright. It is a first-instance decision, an appeal is possible and parallel litigation continues elsewhere.

But it does move the debate beyond the relatively simple question of whether copyright material appeared in a training dataset. Courts are now beginning to ask a more technically demanding question: what did the model actually learn, retain and reproduce?

The answer could ultimately determine how the next generation of AI music services is built, and how the creators whose work helps make those systems possible are paid.