Suno Lost to GEMA. Now Every AI Company With EU Users Has a Problem.

On July 31, a Munich regional court handed down a ruling that should make every AI company selling products to European users pay attention. The court found that Suno, the AI music generation platform, infringed copyrights managed by GEMA, Germany’s music collecting society. Suno had already admitted to training on unlicensed music. The court rejected its defense that training doesn’t require a license.

That part wasn’t entirely surprising. What was surprising was the jurisdictional reasoning. This is the first time a European court has asserted authority over AI training that happened entirely in the United States. The court’s logic: the training data lives inside the model, and the model’s outputs reach German users. That’s enough to make a Massachusetts company’s training runs subject to German law.

If that reasoning holds on appeal, it redraws the map for AI regulation in Europe.

What the court actually ruled

GEMA represents more than 100,000 composers, lyricists, and publishers in Germany. It’s the German equivalent of ASCAP or BMI. The organization filed suit in January 2025, arguing that Suno trained its models on songs in GEMA’s repertoire, including Boney M’s “Daddy Cool,” Lou Bega’s “Mambo No. 5,” and Alphaville’s “Forever Young,” without obtaining licenses or paying songwriters.

The Munich court agreed. It found that Suno unlawfully obtained, processed, and reproduced music represented by GEMA, breaching both German and US copyright law. The court also issued a compelled disclosure order, meaning Suno must provide details about its training data and processes. That disclosure order is significant because it means the specifics of how Suno built its training dataset will eventually become part of the public record.

GEMA’s CEO called the ruling a milestone. That’s not hyperbole. The decision establishes that a European collecting society can go after a US-based AI company in a European court, win, and potentially force licensing negotiations on terms favorable to the rights holders. It also establishes a precedent for damages calculations that could reshape how AI companies value the risk of unlicensed training.

The six songs named in the lawsuit are just the beginning. GEMA’s broader argument is that Suno trained on its entire repertoire, not just these six tracks. The compelled disclosure order could reveal the full scope of what Suno used, which would dramatically expand the damages calculation. If GEMA can prove that Suno trained on thousands of works in its catalog, the financial exposure becomes enormous.

The jurisdiction bombshell

The most consequential part of the ruling isn’t about Suno at all. It’s about where the court says EU copyright law applies.

The court held that German law reaches AI training conducted in the US because the training data was incorporated into the model, and that model serves outputs to users in Germany. Under Section 131 of the VGG, the German statute governing collecting societies, that connection was enough to bring a Massachusetts company’s operations before a Munich judge.

The court went further. It said Suno would have owed GEMA a license even under US copyright law, the law governing where the training physically happened. So the court wasn’t just claiming jurisdiction over the German-facing outputs. It was claiming that the training itself, wherever it occurred, violated copyright.

If this reasoning survives appeal, every AI company with EU users faces exposure. It doesn’t matter that the model was trained in California or Beijing. If the outputs reach European consumers, European courts can hear the case. That’s a fundamental shift from the assumption that AI training happening outside EU borders falls outside EU copyright enforcement.

The ruling also creates a template. GEMA can take this playbook to other EU member states. Denmark’s collecting society, Koda, already sued Suno in November 2025 over similar allegations involving Danish artists. Expect more collecting societies to follow.

What Suno admitted and what it tried to argue

Suno’s defense was straightforward: training on copyrighted data doesn’t require a license because the model learns patterns, not copies. This is the standard AI industry position, and it’s the argument that companies like OpenAI, Google, and Meta have made in various forms.

The Munich court rejected it. The ruling found that Suno’s use of the songs went beyond learning patterns. The platform reproduced identifiable elements of the original works, and the court treated that as infringement regardless of whether the output was a direct copy or a statistical approximation.

Suno also admitted, during the proceedings, that it trained on unlicensed music. That admission made the legal fight harder. When you’ve already conceded the core fact, the argument shifts to whether the law requires a license for that activity. The court said it does.

This is significant because the AI industry has relied heavily on the ambiguity of copyright law as it applies to training. The argument has always been that the law is unclear, that fair use might apply, that training is transformative. The Munich ruling cuts through that ambiguity, at least in Germany. The court said the law is clear enough, and Suno broke it.

For other AI companies watching this case, the takeaway is uncomfortable. The “train first, license later” strategy that many have relied on just got significantly more expensive. The court’s compelled disclosure order means Suno’s training data will eventually become part of the public record. Every company that trained on copyrighted data without a license is now looking at a future where their own training practices could be similarly exposed.

The Warner Music deal and why it doesn’t help now

In November 2025, Warner Music and Suno announced they had settled their lawsuit and signed a licensing deal at the same time. Suno agreed to build new models on Warner’s catalog and acquired Warner’s Songkick platform as part of the package. Universal Music and Sony Music’s lawsuits against Suno remain active, seeking statutory damages that could run into the billions.

The Warner deal was supposed to be a model for how the industry resolves AI copyright disputes. License the data, build new models, everyone moves forward. But the GEMA ruling undercuts that narrative in two ways.

First, the ruling happened after the Warner deal was signed. The court found that Suno infringed GEMA’s copyrights by training on music it didn’t have rights to. The Warner deal only covers Warner’s catalog. It doesn’t address the unlicensed training on GEMA-represented works that the court just found illegal.

Second, the jurisdictional reasoning means that a licensing deal with a US rights holder doesn’t protect an AI company from litigation in European courts. Suno can have a perfectly valid agreement with Warner Music and still lose a lawsuit in Munich over the same training data if that data includes European works.

The precedent this creates is messy. AI companies can’t simply cut deals with the major labels and assume they’re covered. European collecting societies operate independently, and each one can bring its own claims. Suno is now facing litigation from GEMA in Germany, Koda in Denmark, and potentially dozens of other collecting societies across the EU.

What this means for AI companies outside music

The ruling’s implications extend well beyond music generation. Any AI system that trains on copyrighted data and serves outputs to European users is now potentially subject to EU copyright enforcement, regardless of where the training happened.

Consider the implications for AI image generators. Companies like Midjourney and Stability AI trained on billions of images, many of which are copyrighted. If the Munich reasoning holds, European courts could hear cases against these companies even though the training happened in the US. The collecting societies that manage visual arts copyrights in Europe could pursue similar strategies to GEMA.

The same applies to large language models. GPT-style models trained on copyrighted text, copyrighted code, copyrighted datasets of all kinds. If the model serves outputs to European users, a European court can assert jurisdiction. The legal theory is the same: the training data lives in the model, the model serves European users, that’s enough.

This doesn’t mean every AI company will be sued tomorrow. Collecting societies have limited resources and will prioritize cases with the biggest potential impact. But the legal framework just got a lot clearer, and it favors the rights holders. The era of hoping the law is too ambiguous to enforce is ending.

For AI companies, the strategic response is likely to accelerate licensing negotiations. The cost of fighting these cases in European courts, losing, and facing compelled disclosure orders is now higher than the cost of licensing the data upfront. The “train first, license later” approach just got a price tag, and it’s steep. Companies that haven’t started licensing discussions with European rights holders should start now, before the next ruling makes the terms even less favorable.

The licensing roadmap is getting clearer

The GEMA ruling, combined with the Warner Music deal and the ongoing Universal and Sony lawsuits, is sketching out what the future of AI training looks like. It’s a future where licensing is mandatory, not optional, and where the terms are set by rights holders who now have court precedent on their side.

For AI music companies specifically, the path forward involves licensing from collecting societies like GEMA, ASCAP, BMI, and their international equivalents. The Warner deal shows that direct licensing with major labels is possible, but the GEMA ruling shows that it’s not sufficient. You need both.

For AI companies in other domains, the lesson is the same. If your training data includes copyrighted material, and your product serves European users, you need a licensing strategy. Not eventually. Now.

The Munich court didn’t create new law. It applied existing copyright principles to a new technology and concluded that the technology doesn’t get a free pass. That’s a boring conclusion, but it’s the one that matters. The AI industry spent years arguing that training is different, that it’s transformative, that it shouldn’t be bound by old rules. The courts are now saying it is, and it should be.

The licensing roadmap isn’t finished, but the destination is clear. AI companies will pay for the data they train on. The only question is how much, and whether they’ll do it voluntarily or under court order. Suno just found out what the court order version looks like.