A Munich court has handed down the first European ruling against a generative AI music platform, finding that Suno trained its models on copyrighted songs without a licence, stored those works inside the model, and reproduced them through user outputs — and that US fair use does not shield the training, even though the training happened in the United States.
The Munich Regional Court (Landgericht München I) ruled on 31 July 2026 in favour of GEMA, Germany’s performing rights society, in a case that covered six compositions from its repertoire: “Atemlos durch die Nacht,” “Daddy Cool,” “Rasputin,” “Big in Japan,” “Forever Young,” and the refrain of “Mambo No. 5.” The court ordered Suno to cease and desist, disclose its revenue, and pay damages.
💡 THE BOTTOM LINE: The ruling establishes that when an AI model memorises a copyrighted work and can reproduce it on simple prompts, the provider is a direct infringer — not a neutral tool. The court rejected both the EU text-and-data-mining exception and the US fair use doctrine as defences. For AI companies that trained on copyrighted music without licensing, this is a blueprint for liability.
What the Court Found
The court identified four separate acts of infringement, according to the legal analysis by Bird & Bird:
- Reproduction for training in the US — Suno trained on complete recordings obtained from YouTube by stream-ripping, circumventing the platform’s “Rolling Cipher” download restriction. The court applied US copyright law to these acts and rejected fair use.
- Reproduction through memorisation in Germany — the model’s parameters had taken over more than mere statistical patterns; the works themselves were embedded in the weights.
- Communication to the public through offering the model — making a model that holds memorised works available to users is itself an infringing act, regardless of whether anyone actually retrieves the work.
- Reproduction and communication through outputs — the generated tracks reproduced the melodies of the original compositions.
The court found that GEMA’s test prompts were simple and open-ended — lyrics, style, and title only — leaving melody, harmony, rhythm, and tempo to the model. When the output nevertheless reproduced the original melody, the court concluded the music could only have come from the model itself, not from the user’s prompt.
Suno argued that its model holds only trained parameters, not the recordings themselves, and that any resemblance was the user’s doing. The court disagreed, citing memorisation research linking retention to how often a work appears in the training set — a mechanism Suno itself had described for popular songs.
Fair Use Rejected — And a Warning About TDM Exceptions
The court’s fair use ruling may be the most consequential part of the decision. Suno trained its models in the United States and argued that the training fell under the US fair use doctrine. The Munich court disagreed, applying US copyright law to the American training acts and finding that fair use did not apply.
On the EU side, the court examined the text and data mining (TDM) exception under Section 44b of the German Copyright Act. It found that while the exception covers the copies made during analysis — format conversion, backups, assembling the training corpus — it does not cover reproductions inside the model itself. Memorisation goes beyond analysis: the works are not merely evaluated but taken over into the parameters. The premise the exception rests on — that analysis leaves the author’s own exploitation untouched — no longer holds.
The court added a point with wide reach: if current technology cannot prevent memorisation, training on protected works is entirely excluded from the exception. A business model that helps itself to the intellectual property of others free of charge is, in the court’s words, unknown to both EU and German law.
The court also found that Suno’s access was unlawful. The six works could be listened to on YouTube, but the Rolling Cipher exists to prevent downloading. By circumventing it, Suno obtained access in violation of Section 95a of the German Copyright Act, which protects technical protection measures.
Suno separately argued that compliance with the EU AI Act could serve as a justification. The court rejected this: AI Act compliance and copyright compliance are strictly separate. The training-data summary under Article 53(1)(d) exists to help rightsholders enforce their rights, not to replace licences. Suno’s reading, the court noted, would turn a transparency duty into an immunity.
The Provider Is Responsible, Not the User
A key question in AI copyright cases is whether the provider or the user bears responsibility for infringing outputs. The Munich court placed that responsibility squarely on Suno.
The court found that Suno acted as a direct infringer, not an intermediary. Suno offers both the model and the application together and created the infringing outputs itself on simple, open-ended prompts. The outputs count as Suno’s own information rather than third-party content it merely conveys.
This is distinct from a platform that hosts user-generated content, which would be assessed under the intermediary framework with a “central role” and “intent” test. The court’s reasoning leaves a door open for providers whose models do not memorise works, or whose infringing outputs appear only on steering prompts — but the risk of memorisation falls entirely in the provider’s sphere.
Suno’s Response and the Broader Landscape
Suno said it disagrees with the ruling and is evaluating all available options, including an appeal to the Munich Court of Appeal. In a statement to Variety, a Suno representative said the ruling “rests on a fundamental mischaracterization of how Suno’s technology works, how it is used and how U.S. law applies.”
The company, valued at $5.4 billion after a $400 million Series D round, has faced parallel litigation in the US from Universal Music Group and Sony Music. Warner Music Group reached a licensing agreement with Suno last November, which also included Suno’s acquisition of the concert ticketing platform Songkick — a deal we covered here.
The GEMA ruling builds on a November 2025 decision in which the same court found OpenAI had trained its models on unlicensed lyrics. GEMA was awarded unspecified damages in that case. The society has since released a proposed licensing system for AI companies’ use of music.
The broader context includes the Anthropic $1.5 billion copyright settlement — the largest in American history — where the judge found training was fair use but the company had pirated the works it trained on. A report from 404 Media indicated Suno may have similarly scraped material from YouTube without permission.
What This Means for New Zealand
APRA AMCOS, which represents 128,000 songwriters, composers, and music publishers across Australia and Aotearoa New Zealand, welcomed the ruling. CEO Dean Ormston said: “Munich has called it what it is. Not innovation. Theft.”
Ormston named specific New Zealand artists — Midnight Oil, Cold Chisel, Sia, Crowded House, Lorde — whose works have been found inside AI systems without consent or payment. “This is not just a German case. It’s a warning to every AI developer, in every territory, including this one. You cannot train on the world’s music and argue the world’s laws don’t apply to you.”
New Zealand’s copyright framework does not currently have a text and data mining exception comparable to the EU’s. The question of whether AI training on copyrighted material without a licence constitutes infringement under NZ law remains untested in court. The GEMA ruling is likely to inform how New Zealand courts and policymakers approach the same question — particularly the finding that offering a model which has memorised a work is itself an infringing act, separate from any user output.
The EU AI Act’s Phase 2 enforcement, which began on 2 August, adds another layer. While the Munich court was explicit that AI Act compliance does not equal copyright compliance, the transparency obligations under Article 53 — which require AI providers to publish summaries of their training data — give rightsholders the information they need to identify unlicensed use.
❓ FAQ
Can Suno appeal the ruling? Yes. Suno can appeal to the Munich Court of Appeal. The ruling is not final, and Suno has stated it is evaluating all available options.
Does this ruling apply outside Germany? The ruling applies to Suno’s activities in Germany, but the court’s reasoning — particularly its rejection of fair use for training conducted in the US — is likely to be cited in copyright cases in other jurisdictions. Courts in other countries are not bound by it, but the legal analysis may be persuasive.
What is “memorisation” in an AI model? The court defined it as a model whose parameters take over more than the patterns and correlations it normally extracts — at least part of the content itself. An exact copy is not required. What matters is not how much of a work is taken over, but what is taken over.
How does this differ from the Anthropic settlement? The Anthropic case was settled in a US court for $1.5 billion. The judge found that Anthropic’s training on authors’ work was fair use but that the company had pirated the works it trained on. The GEMA ruling is a court decision (not a settlement), it rejected fair use entirely, and it found that the act of offering a memorised model is itself infringement.
🔍 THE BOTTOM LINE
The Munich ruling is the first court decision anywhere to find a generative AI music company liable for copyright infringement across training, memorisation, model offering, and output generation. The court’s rejection of fair use — applying US law to US training — and its finding that the TDM exception does not cover memorisation give rightsholders a legal framework that did not exist a week ago. Suno will appeal, and the legal battle is far from over. But for AI companies that trained on copyrighted material without licensing, the question is no longer whether they can be sued. It is how many courts will reach the same conclusion.
