Munich court finds Suno liable for training and outputs
Munich Regional Court I ruled that Suno infringed copyrights in both training and generated music, rejecting the company's US-style fair use defence. Suno says it disagrees and is weighing an appeal.
Munich Regional Court I ruled on 31 July that AI music generator Suno infringed copyrights held by German collecting society GEMA, finding the company liable both for how it trained its models and for what those models produce.
What the court found
The court concluded that Suno's models had memorised six songs in GEMA's catalogue, including works associated with Kristina Bach and Frank Farian. When prompted with a title, a style tag and lyrics — but no explicit musical instructions — the system returned output the court considered substantially similar to protected elements of the originals. Liability was placed on Suno rather than its users: the company chose the training material and designed the architecture that led to memorisation.
Suno had argued a US fair use defence. The court applied US law to that argument and rejected it, finding every factor weighed against the company, and distinguishing the case from earlier American proceedings on the grounds that here the training material resurfaced in outputs available to users. Court filings also described Suno using stream-ripping tools to pull audio from YouTube, circumventing the platform's rolling-cipher protection — a separate act the ruling treats as unlawful independent of the copyright question.
Suno responded that it disagrees with the ruling, saying it rests on a flawed understanding of its technology, of how the platform is actually used and of how American law applies, and that it is evaluating all available options, including an appeal. GEMA chief executive Tobias Holzmüller said the decision showed that AI models built on stolen intellectual property enjoy no legal protection and that providers must license the works they use. The judgment is not final.
Why it matters
This is the first substantive court ruling to hold a generative music company liable on both ends of the pipeline at once — inputs and outputs — and it does so in a jurisdiction with no fair use doctrine, while nonetheless walking through the US test and finding it fails. For music generators, the memorisation finding is the sharper problem: it converts a research-grade concern about training-data leakage into a concrete liability standard that a plaintiff can demonstrate with a handful of prompts. The stream-ripping detail compounds it, because circumvention claims do not depend on how a court views model training at all.
The practical consequence is that any music model shipped in Germany now faces a defined test, and rights holders elsewhere in Europe have a template. An appeal, which Suno says it is weighing, would decide how durable the reasoning proves, but it will be cited well beyond Munich — including by the labels currently pressing similar claims in the United States.