What the court decided on 31 July
The Munich Regional Court I largely granted GEMA's claims against AI music generator Suno for injunctive relief, disclosure, and damages. The court didn't just assess whether a specific AI-generated output was too similar to a protected work - it evaluated the entire interplay across three phases. During training, Suno reportedly reproduced protected songs without authorization, including through stream-ripping from YouTube while circumventing technical protection measures. In model storage, the court found the works weren't merely learned as abstract stylistic patterns, but reproducibly 'memorized' within the model's parameters - Suno had argued the model only extracted generalized features, which the court rejected. In output generation, the court found that generated music containing protected elements recognizable to an average listener constitutes an independent copyright infringement.
Six well-known works from GEMA's repertoire are affected: 'Forever Young', 'Atemlos durch die Nacht', 'Mambo No. 5', 'Rasputin', 'Big in Japan', and 'Daddy Cool'. GEMA CEO Dr. Tobias Holzmüller commented: 'AI models based on the theft of intellectual property are not protected by the legal system.' GEMA board chair Dr. Ralf Weigand added: 'Creativity has value, and creators' rights apply in the age of artificial intelligence too.' As things stand, the ruling isn't yet final.
Why liability currently sits with the provider
For the specific outputs generated, the court assigns responsibility primarily to Suno as the provider, not to the individual users who generated a song. The reasoning: users typically formulated only simple, open-ended prompts, without giving detailed musical instructions. The court locates 'control over the act' (Tatherrschaft) - meaning decisive control over the specific copyright infringement - with Suno, because the model's architecture and the works stored within it sit in the provider's sphere, regardless of the specific prompt a user entered.
The part of the reasoning that reaches beyond music
This is exactly where the point sits that's missing from coverage framed mostly as a music-law story: the court explicitly leaves open that this assignment of responsibility can reverse. If a user works with repeatedly adjusted, evaluative, or steering instructions aimed at a specific protected work - not just entering 'make me a pop song', but iteratively fine-tuning until the result increasingly resembles a specific original - control over the act potentially shifts to the user, per the ruling's logic.
This distinction isn't specific to music. Iteratively refining a prompt against a specific reference - 'make the logo more like this well-known brand', 'write the text in the style of this specific author', 'generate an image closer to this specific artwork' - is an everyday working method in marketing and content teams using generative AI for text, images, or audio. Per this ruling's logic, that exact behavior is the point at which responsibility could shift from provider to user - even though the court didn't have to conclusively decide this question for the case at hand, because the prompts actually used were found to be open-ended and generic.
Why this matters already, despite not being final
The ruling isn't yet final; an appeal is possible. That doesn't change the fact that the legal reasoning already offers guidance - particularly for a company deciding today how to use generative AI in marketing, content production, or product development. A German court has, for the first time, explained in detail the conditions under which responsibility for AI-generated content shifts from provider to user - and that condition is tied to behavior that's routine practice for many content teams: deliberately, iteratively refining a prompt against a specific reference.
That doesn't mean every instance of iterative prompt refinement automatically constitutes copyright infringement, or that every company using generative AI carries an acute risk. It means the blanket assumption 'liability sits with the AI vendor anyway, not with us' no longer holds unconditionally after this ruling. Exactly where the line runs between generic prompting and targeted, steering prompting will only become clear through further proceedings - until then, caution is the sensible response, not certainty.
What this means in practice
Legal commentary on this ruling consistently converges on one central recommendation: companies shouldn't rely solely on an AI vendor's assurances, but should take their own precautions.
- For commercially used AI-generated content (text, images, music), document whether and how a prompt was iteratively refined with reference to a specific, well-known work or brand - per the Munich ruling, exactly this pattern is the point at which liability could shift.
- Set internal guidelines for content teams that explicitly discourage prompts deliberately steering toward a specific protected work or brand ('make it more like X'), in favor of generic, open-ended phrasing.
- Critically review AI vendors' contractual assurances about legal compliance rather than treating them as complete liability protection - the Munich ruling shows that your own usage pattern can help determine how liability is allocated, regardless of what a vendor contractually promises.
- Keep an eye on how the case develops (appeal, possible follow-on rulings on text or image generators), since the exact boundary between unproblematic and risky prompting will only be sharpened by future rulings.
The real value of this analysis isn't a warning about Suno or music generators specifically, but the transfer of a legal principle: a German court has, for the first time, precisely described that the manner of prompt use itself can be liability-relevant - a standard that isn't limited to music, but applies to any form of generative AI a company uses commercially.