GEMA didn't hand Suno a recording of "Mambo No. 5" and ask for a remix.
Its testers typed in the original lyrics, the song title and a description of the musical style. They didn't supply the melody, the harmony, the rhythm or the arrangement. Suno produced music containing recognizable elements of the original composition anyway.
That detail sits at the center of a decision issued by the Munich Regional Court on July 31, 2026. The court largely upheld GEMA's claims against the American AI-music company, finding infringement tied to the training of Suno's models, the storage of protected material inside those models, and the music generated for users in Germany.
The judgment covers six compositions: "Atemlos durch die Nacht," "Rasputin," "Big in Japan," "Forever Young," the refrain of "Mambo No. 5," and "Daddy Cool." Lyrics weren't part of the infringement claims decided here. Suno has said it disagrees and is considering an appeal — the decision isn't final.
Even with those qualifications, this is a serious result. A European court examined an AI-music model rather than a text generator, found protected compositions reproducible inside it, and rejected Suno's fair-use defense for training performed in the United States. For songwriters, publishers, producers and artists using generative tools, the reasoning matters more than the headline.
The court found infringement at three separate stages
Arguments about AI training tend to collapse everything into one question: can a company train on copyrighted material without permission? The Munich court refused to flatten it that way and identified several distinct uses.
First, it found that copies of the six compositions were made during training in the United States. Per the court's official summary, Suno's training dataset contained the works, extracted from YouTube using stream-ripping that bypassed a technical download protection.
Second, it concluded the compositions were memorized inside versions 3.5 and 4 of Suno's models stored on German servers. Memorization here means the model retained enough protected expression that it could be recovered through generated output. Given the length and complexity of the musical similarities, the court found coincidence implausible.
Third, the resulting outputs reproduced and communicated recognizable original elements in Germany.
That sequence is the thing to hold onto:
training copy → model memorization → infringing output
It matters because a future case might land differently if a claimant proves only the first stage but can't show that protected expression survived inside the model or surfaced in its output. The decision does not say every act of machine learning is automatically infringement — it deals with models from which substantial musical similarities could actually be produced.
The prompts didn't supply the music
Suno argued that GEMA's testers manufactured the outputs through targeted, repeatedly refined prompts. The court didn't buy that characterization.
GEMA entered the lyrics, title and intended style — but never instructed the system to use the original melody, harmony, rhythm or arrangement. The court described the prompts as simple and open-ended, and held Suno responsible because the company selected the training material, operated the models, and controlled the architecture that let the compositions be memorized in the first place.
That's a real problem for a defense AI companies lean on constantly: the generator is just a tool, and the user decides what it makes. Sometimes that's true. Here the court concluded the model itself determined the infringing musical content. Think about what actually happened — a user asked for a song with words and a broad stylistic direction, and the missing music had to come from somewhere. When the resulting melody, harmony and rhythm turned out substantially similar to a work sitting in the training data, the court attributed that to Suno's system, not to musical instructions the user never gave. Expect this to be one of the most closely watched parts of any appeal.
Text-and-data-mining rules didn't cover memorization
European copyright law includes exceptions permitting certain text and data mining, and those provisions get floated regularly as a legal foundation for AI training — subject to their conditions and any rightsholder reservation. The court found the German TDM exception didn't cover what happened here, and its reasoning turned on memorization: the models hadn't merely examined the works to learn general patterns and then left the protected expression behind. Content from the training material, the court said, was incorporated into the model parameters in a form that could be reproduced.
That distinction could end up being the most consequential part of the whole case. A model that learns broad relationships between tempo, instrumentation and genre without returning recognizable passages raises one legal question. A model that can regenerate substantial parts of a specific protected composition raises a very different one.
A caution for independent musicians on the word style: copyright generally doesn't hand anyone ownership of a whole genre, mood or production technique. This case was about recognizable musical expression from named compositions — not the abstract idea of making something that sounds like 1980s synth-pop. That line matters as much when you're defending your own work as when someone accuses you.
The court also applied US fair-use law
The training happened in the United States, so Suno argued the German court had no authority over the American part of the dispute and that the copies were protected by US fair use. The court rejected both arguments — finding, under German rules governing collecting societies, that it had jurisdiction over related claims involving acts in the US, and then applying American copyright law to that training activity.
Per the court's summary, all four fair-use factors weighed against Suno. The judges distinguished the case from the US book decisions Bartz and Kadrey, where AI training had been treated as fair use under particular circumstances — in those, the training material wasn't made substantially available to users through the output. Suno's models, by contrast, generated music materially similar to the originals after comparatively simple prompts.
This is an ambitious stretch for a first-instance German judgment, not a final statement of US law, and American courts hearing other Suno lawsuits aren't required to follow it. But it does undercut the idea that an AI company can say the words "fair use" and end the conversation. What the finished model does — and what users can pull back out of it — is now part of the analysis.
What GEMA actually won
The court largely granted GEMA's requests for an order preventing the unlicensed uses at issue, information about the relevant activity and revenue, and damages with the amount still to be determined.
GEMA frames the decision as confirmation that AI providers must license and pay for the protected works they exploit — which is the successful claimant's interpretation, and licensing is plainly the commercial outcome it's been chasing. But the ruling itself doesn't create a universal AI royalty or set a payment rate for training music models. It determines liability involving six compositions and opens the path toward damages and licensing negotiations. It also doesn't automatically put money in the pocket of every GEMA member whose music may have entered Suno's datasets; other works would still have to be identified and tied to an enforceable claim or a broader settlement. Collection societies now have a useful legal and evidentiary template. They don't yet have an automatic royalty pipeline.
What Suno says
Suno told Music Business Worldwide that it built its models to create new music rather than reproduce existing songs, and that it included safeguards in the platform. It said the judgment mischaracterized its technology, its use, and the application of US law, and that it's weighing its options including an appeal.
That's worth holding in view, because the public court release presents the judge's findings while GEMA's announcement presents the claimant's victory. The full written judgment and any appeal could surface further technical disputes about model storage, similarity testing and causation. For now the ruling is enforceable only under the applicable German procedural rules and is explicitly described by the court as not legally final — so anyone summarizing it as "Suno training declared illegal worldwide" has left the actual decision behind.
What it means for songwriters
You can't assume this judgment has identified or monetized any use of your catalog. What you can do is make future enforcement easier.
Keep authoritative copies of your compositions, demos, stems, scores and release masters. Preserve creation dates and collaborator agreements. Register works accurately with the appropriate society, publisher or administrator, and keep titles, writer names, shares, IPIs and recording identifiers consistent.
If a suspicious output turns up, save more than a streaming link. Record the complete prompt and account used; the model and version; the date, time and territory; every generated variation; the original downloaded file; screenshots or a screen recording of the process; a musical comparison prepared by a qualified professional; and proof of ownership or representation for the earlier work.
And a discipline point: don't publicly accuse a company or a user of theft just because a track shares a genre, a chord progression or a vocal texture. Copyright analysis turns on protected expression, access, similarity and the applicable law — GEMA's evidence was specific enough for a court to find recognizable melody, harmony and rhythm across named works. "This feels like my song" isn't that.
What it means for artists using Suno or other generators
The judgment did not declare every Suno output infringing, and it didn't rule that you commit infringement merely by opening the app. The risk shows up when generated material reproduces protected expression.
Before you commercially release an AI-generated or AI-assisted track:
- Listen past the surface. Check the central melody, hook, harmony, rhythmic phrasing and arrangement against music you know.
- Search distinctive lyric and title phrases. Lyrics were outside this case, but they carry their own copyright exposure.
- Keep the prompt history and generated alternatives. If a dispute arises, that trail may help show what you supplied and what the model introduced.
- Review the tool's commercial terms. They may let you use an output without guaranteeing it's free of third-party claims.
- Don't prompt for a near-copy of a named song, a living artist or a recognizable vocal identity.
- Replace questionable elements with genuinely original human performance or composition before release.
- Be accurate in the AI disclosures you deliver through your distributor.
A subscription is not copyright insurance. "Commercial use allowed" usually describes the relationship between the AI provider and its customer — it can't grant rights the provider doesn't hold. And one more wrinkle: although the Munich court placed responsibility on Suno for these outputs, that's no promise of immunity for a user who distributes, monetizes or publicly performs infringing material somewhere else.
What labels and distributors should change
The immediate operational lesson is provenance. A generic contract warranty saying "the client owns all rights" isn't much use when neither party knows how the recording was actually made. The goal is enough information to evaluate risk without demanding every private studio file.
A proportionate intake process can ask:
- Was generative AI used in the sound recording or the composition?
- Which service and model produced the material?
- Did it generate the lead vocal, melody or primary instrumentation?
- Does the output imitate an identifiable artist or existing work?
- Has the client reviewed the provider's commercial terms?
- Can the client supply prompts and project evidence if challenged?
- Are all required AI labels included?
- Is there a process to suspend delivery while a credible claim is reviewed?
The point isn't to reject every release involving AI. It's to avoid learning the entire production history only after a takedown lands. High-risk content deserves human review — an AI-generated lead vocal resembling a famous performer warrants far more scrutiny than machine-assisted noise removal on a human recording.
What the ruling does not settle
This decision leaves the big questions open. It doesn't determine whether training is infringing when no protected material can be recovered from the outputs. It doesn't bind courts outside Germany. It doesn't fix the damages Suno will ultimately pay, and it doesn't establish a standard license rate. It may change on appeal. It also concerns compositions rather than ownership of the underlying commercial masters — and GEMA's testers used original lyrics, titles and style instructions, a factual pattern that may differ from ordinary prompts. Future cases will test different models, datasets, safeguards and outputs.
Even so, the boundary is clearer than it was. A court was willing to look inside the full chain — copied training music, to model parameters, to the generated track — and treat each stage as legally relevant. AI companies can no longer assume that moving training abroad, calling the model transformative and blaming the prompt will always be enough.
For creators, this isn't a royalty check yet. It's the thing that usually comes first: a court recognizing that the music inside a machine didn't arrive there by magic.
FAQ
Did the court rule that all AI music is illegal?
No. It found infringement involving six specific compositions across training, model memorization and outputs. It explicitly does not hold that every act of machine learning, or every AI-generated song, is infringing.
Is it now illegal to use Suno?
The judgment didn't rule that using the app is infringement. The legal risk attaches to outputs that reproduce protected expression from existing works — which is why checking your output and keeping your prompt history matters before release.
Does this ruling apply in the United States?
It's a first-instance German decision. It doesn't bind US courts, even though the court applied US fair-use law to the American training activity. Other Suno lawsuits in the US aren't required to follow it, and Suno may appeal.
Does "commercial use allowed" in an AI tool protect me?
Not from third-party claims. That language usually describes your relationship with the provider; it can't grant rights the provider doesn't hold, and it doesn't guarantee the output is free of others' copyright.
Does this mean songwriters get paid now?
Not automatically. The ruling opens the path to damages and licensing negotiations for the works at issue; it doesn't create a universal AI royalty or a payment pipeline for every affected writer.
I released a track partly made with AI. What should I do?
Review the central melody, hook and arrangement against known works, keep your prompt history and generated alternatives, confirm the provider's terms, replace anything that resembles a specific existing work, and disclose AI use accurately through your distributor.
Keep your releases provable
However the AI-copyright picture develops, the artists who stay safest are the ones who can show how a record was made. That means accurate metadata, consistent identifiers, and provenance you can actually produce when someone asks.
CREWPORT validates your metadata before delivery and keeps your identifiers, credits and release data together — so what you declare matches what you made.
This article explains a court decision for a general audience; it isn't legal advice. The ruling is a non-final, first-instance judgment and may change on appeal — for anything affecting your own catalog, get advice from a qualified lawyer.
Sources
- Munich Regional Court — Official Summary of the GEMA v. Suno Judgment, July 31, 2026
- GEMA — Statement on Its Victory Against Suno, July 31, 2026
- Reuters — German Court Rules Suno Violated Copyright, July 31, 2026
- Music Business Worldwide — Suno Responds to German Copyright Judgment, July 31, 2026