German Court Rules Against Suno: Who Is Liable for AI Music in Your Ads

Suno lost. Its users did not.

A court in Munich published a decision on a Friday morning, and by the afternoon the question was circling every marketing team that has ever dropped a generated track under a product cut. Are we the ones on the hook now? The short answer is no. The court went out of its way to say so, in a sentence that most of the coverage skipped past.

On July 31, 2026, the 42nd Civil Chamber of the Landgericht München I, the Munich Regional Court I, largely granted GEMA's claims against Suno for injunctive relief, disclosure and damages in case 42 O 763/25. The chamber found that six well-known musical works are reproducibly contained in Suno's v3.5 and v4 models and recognizable in what those models produced. It rejected a US fair-use defence. It rejected Germany's text-and-data-mining exception. And it put responsibility for the infringing outputs on Suno rather than on the people typing the prompts.

Most of the coverage of that judgment is written for the music industry. This is written for the person who has to decide, this week, whether the soundtrack layer of a live campaign is a problem. What follows is what the court actually held, why the liability landed where it did, what the ruling does not say, and where the real exposure sits for someone buying media. It is journalism about a judgment, not legal advice.

What the Munich court actually decided

A first-instance judgment, six songs, three remedies

GEMA filed against Suno at the Landgericht München on January 21, 2025. It is a collecting society, representing composers, lyricists and music publishers, so the rights at issue were the musical compositions, not the sound recordings. Infringement through the lyrics was expressly not part of the dispute. The case narrowed to six well-known works:

• "Atemlos durch die Nacht", by Kristina Bach.
• "Rasputin", by Frank Farian, Fred Jay and George Reyam.
• "Daddy Cool", by Frank Farian.
• "Big in Japan", by Marian Gold, Bernhard Lloyd and Frank Mertens.
• "Forever Young", by the same three writers.
• The refrain of "Mambo No. 5", by David Lubega and Christian Pletschacher.

The chamber largely granted all three of the claims GEMA brought, and each does different work:

• Injunctive relief. A stop order against continuing the conduct held unlawful.
• Disclosure. An obligation to hand over the information needed to calculate what is owed.
• Damages. An amount that follows from the disclosure, so it is not yet public and not yet fixed.

The disclosure limb is the one people skim and shouldn't. Until it is satisfied, nobody outside the case knows the size of the eventual bill, which is also why headlines about this ruling carry no number.

What memorization meant here, as evidence

The court was convinced that the six works are reproducibly contained in the defendant's models, versions v3.5 and v4, and noted that those models were stored on servers in Germany. It described memorization the way information-technology research does: a model may not merely extract information from a training set, but end up with the content itself carried into the parameters specified after training.

That was proven, in the chamber's view, by comparing the works in the training data against the playback in the outputs. Given the complexity and length of the pieces, it ruled out coincidence as the cause. The legal consequence is the part worth holding onto: the reproduction the court found sits inside the model, as an infringement of the reproduction right under section 16 UrhG, before any user shows up at all.

The prompt that produced the evidence

GEMA's test method matters, and it comes straight from the court's own release. For each song, GEMA entered the original lyrics, the desired musical style, and the title of the work into the generator's input box. The prompts contained no specification of melody, harmony, rhythm or arrangement. Suno produced results in which the court recognized the original elements of the source tracks anyway.

Keep that detail. It does real work later in this post, and it is the single most contestable thing in the judgment.

Why responsibility landed on Suno, not on the person prompting

Suno's argument: the user broke the chain

Suno's defence had a structure worth understanding, because every generative tool will run some version of it. The company argued the works were neither contained nor stored in the model at all, and that its weights and parameters represent mathematically learned patterns and generalized features drawn from training data. Any resemblance in the outputs, on that account, came from prompt-induced narrowing of the search space plus statistically learned patterns.

Then the attribution argument: the outputs were the result of complex, iteratively refined prompts written by the plaintiff, so the deliberate, targeted user action interrupted the causal link between the model and what it produced. In plain terms, if the output looks like "Rasputin", that is because someone worked hard to make it look like "Rasputin".

The court's answer: whoever built the model determined the output

The chamber rejected it in one line: for this, the defendant and not the users are responsible. The reasoning behind that line is what will get quoted for years. The prompts were simple and open-ended. Suno operates the models. Suno selected those musical works as training data and trained on them. Suno is responsible for the architecture of the models and for the memorization of the training data. Therefore the models substantively determined the outputs.

The court went one step further than the outputs. It held that merely offering the model, and the defendant's own application for generating music, already infringes the unnamed right of communication to the public under section 15(2) UrhG. That is a finding about the product existing, not about any particular file it made.

What that does, and does not, do for you

This is the most useful passage in the judgment for anyone buying media, and it is being reported backwards in a lot of places. A court looked at the whole generation chain and put the duty at the point where the training data was chosen. It did not invent advertiser liability. It did close to the opposite: it named the model operator as the responsible party for outputs that reproduce protected works.

What it does not do is clear your ad. The judgment resolves who infringed by making and serving those specific outputs. It licenses nothing on your behalf, it decides no claim about your campaign, and it has nothing to say about the separate agreements you signed with the platforms that run your creative. Those are different documents enforced by different people on a much faster clock.

The part of the ruling that travels

A European court reached acts of training in the United States

Under section 131 VGG, the statute governing how collecting societies exercise copyright, the chamber held it had international jurisdiction over the claims arising from infringing acts on US territory. GEMA describes this as the first time a court in Europe has decided on the training of these systems in the USA, and that framing is roughly why the ruling is being called a landmark rather than a local dispute.

The German side of the case had its own hook. The models were on servers in Germany, which is what put the reproduction inside the model within reach of German copyright law regardless of where the training happened.

US law applied, and fair use failed

Under the country-of-protection principle, the court applied US copyright law to the copies made during training in the United States, and concluded they were not covered by the fair-use doctrine in 17 U.S.C. section 107. It distinguished the US Bartz and Kadrey proceedings, where American courts had treated AI training as fair use, on a specific ground: in those cases the training data was not, or not substantially, made accessible to users in the outputs. Here, simple and open-ended inputs produced outputs substantially similar to the originals.

All of the factors to be examined under the fair-use test, following the US Supreme Court's Warhol decision, were found to weigh against Suno. Not most. All of them.

The European exception did not save it either

Suno also relied on the text-and-data-mining exception in section 44b UrhG, the provision that permits automated analysis of lawfully accessible content. The chamber held that the reproduction inside the models is not covered by it. The line it drew is between analysing content and retaining it, and the retention finding is what put the case outside the exception.

There is one more fact in the release that will matter more over time than any of the above. The court records that Suno used stream-ripping techniques to extract the works from YouTube, bypassing the Rolling Cipher, a technical protection measure meant to prevent downloading. How training material was acquired is a separate question from whether training on it is fair, and it is the question that survives most fair-use outcomes. That is the same fault line running through the Sony lawsuit against Udio, where the anti-circumvention count is the durable one.

What the ruling does not say

Two columns are worth keeping straight, because almost every summary of this judgment blurs them:

What the judgment decided	What it did not decide
Suno infringed six GEMA-represented works	Whether your ad is cleared
Memorization inside v3.5 and v4	That all model training is unlawful
Fair use failed on every Warhol factor	How US courts will rule elsewhere
Section 44b does not cover retention	Where lawful analysis ends
Suno is responsible, not users	Any advertiser liability
Not final, appeal open	Settled European law

It is not "AI training on music is now illegal in Germany"

The wide version of this story is the one that will spread, so it is worth being exact. The infringement found here is the reproduction of six identified works retained in two identified model versions, plus the outputs that reproduced recognizable elements of them. Memorization was proven in this case, on this evidence. That is a factual finding about particular songs in particular models, not a general holding that training is unlawful.

The distinction also travels less far than it sounds. The court is not saying music models memorize and text models do not. It is saying reproduction was concretely proven here and was not proven in the American cases it distinguished.

The prompts were not normal prompts

Here is the part I would push back on, and it is a description rather than a legal conclusion. The chamber characterized GEMA's prompts as simple and open-ended. Those prompts contained the complete original lyrics, the musical style, and the song title. Nothing about melody was specified, which is the court's point, but the identity of the target song was specified about as precisely as a text box allows.

Nobody producing an original bed for a skincare ad prompts that way. Whether a targeted prompt designed to surface protected content represents ordinary use of a generative system or a special case is one of the genuinely open questions in this field, and it is being fought over in other courtrooms right now. A judgment that rests partly on calling those prompts ordinary has an exposed edge on appeal.

And it is not final

The court says so plainly: the judgment is not final. This is a first-instance decision from one chamber. Suno's response was that it disagrees, that from the beginning it trained its models to create new songs rather than reproduce existing ones, and that it is "evaluating all available options, including an appeal." GEMA's earlier November 2025 win against OpenAI at the same court is already on appeal at the Oberlandesgericht München, which tells you the shape of what comes next.

So write it down as a data point with a strong direction of travel, not as settled law. Anyone telling you the question is now closed in Europe is selling something.

The risk that actually reaches your campaign

An injunction against your supplier is a supply-chain event

Here is the mechanism nobody puts in the headline. An injunction, a disclosure order and a damages claim all run against the tool maker. Nothing legal happens to you. What happens to you is that the tool changes, in ways that are familiar to anyone who has lost a model mid-campaign:

• Output filters that refuse prompts they accepted last week.
• Catalogue or style restrictions that quietly narrow what the generator will make.
• Regional availability changes, since the German server finding is a geography argument.
• Pricing moves, because licensing costs land somewhere.

The numbers make it concrete. Take a team running 20 live ads a week with one house music bed reused across variants. If that bed has to be swapped because the generator behind it restricts what it will produce, the swap is not one file. It is every live variant carrying that audio, plus the cuts sitting in review, plus the versions already uploaded to two ad accounts. One asset, twenty rebuilds, and none of them were on the roadmap. This is the same failure shape as a model deprecation, and it is why creative operations exists as a discipline: the cost is never the asset, it is the fan-out. The size of that fan-out is a number you can decide deliberately rather than discover during an incident, which is the argument in how many ad creatives you actually need.

The platform contract is the one that bites first

Whatever a German court decides about a US model maker, the agreement that gets enforced against your account is your ad platform's. Those rules move at review speed, not litigation speed, and they are already specific about machine-generated creative: what has to be labeled as AI-generated, what TikTok requires you to disclose, and how Google treats AI-generated imagery in ads. A rejected ad on Tuesday costs a campaign more than a foreign judgment does.

The question you will be asked is provenance, not doctrine

When this comes up in your world it will not arrive as a legal theory. It arrives in one of three shapes, and none of them mention Munich:

• A brand-safety questionnaire from a client or an agency partner.
• A platform review flagging an audio match on a live creative.
• A rightsholder or a label contacting the brand name on the ad.

The answer is either a source you can name or a shrug, and only one of those is survivable. If you want the practical version of that layer, the soundtrack clearance paths for AI ads piece walks the options, and the state of AI music models for ads covers what the generators can and cannot do today. The pattern generalizes past audio, too, which is the lesson from the copyright suit over an AI ad generator's meme template: the risky asset is rarely the one you thought about. It is also getting harder to keep the audio layer separate enough to answer for: video models increasingly generate the soundtrack in the same pass as the picture, which is exactly what MiniMax H3 ships with native stereo audio does. When the music never existed as a file you chose, "where did this come from" gets a worse answer.

How Novoads keeps the soundtrack a decision you can document

Novoads generates the layers a creator-style ad is actually built from: an AI actor delivering your script, the voiceover, captions, and ad images made from a product photo you upload. Music is not one of those layers, which means the bed never arrives bundled inside a render you cannot account for. It stays an explicit choice you make in the edit, from a source you can name, on an asset you can point at months later.

That is also how the best creator-style ads already sound, since a voice and room tone read as a recommendation and a polished bed reads as a commercial. You can try the actor and script flow for $1: $1 for 3 days of access, cancel anytime.

Provenance is becoming a property of the tool, not just the track

The Munich chamber did something more interesting than picking a winner. It looked at a generated file, asked who decided what came out of it, and answered: the party that chose the training data and built the model. That is a supply-chain question wearing legal clothes.

It is also the question everyone downstream is converging on. Not "is this legal", which nobody can answer for you today, but "whose decisions are inside this file, and can you name them". Teams that can answer that for every second of audio and every frame of video they ship will find this whole news cycle uneventful. Everyone else gets to answer it under a deadline, in front of a client, about a campaign that is already running.