A German court just ruled that training an AI on music is copying it — and that doing it in America is no defence.

The reported version of what happened in Munich last Friday is that a German court "banned" an American AI music app, or "ruled against AI," or "sided with the music industry." All of that is roughly true, and none of it tells you what the judgment binds anyone to do. On 31 July, the Regional Court of Munich I decided, in case 42 O 763/25, that Suno — the US company whose model generates finished songs from a line of text — had infringed the copyrights administered by GEMA, Germany's music collecting society. The useful questions are the ones the headline skips. What exactly did the court find infringing? On what statutory basis? And why should a ruling from a civil chamber in Bavaria matter to a company that trained its model on servers in another hemisphere?

Start with what was actually in front of the court, because the court was unusually concrete about it. GEMA did not argue in the abstract about the nature of machine learning. It put six songs on the table — among them "Atemlos durch die Nacht," "Forever Young," "Big in Japan," "Mambo No. 5," "Daddy Cool" and "Rasputin" — and, by accounts of the hearing, played the originals and Suno's generated outputs back to back in the courtroom. The chamber, presided over by Judge Elke Schwager, found the resemblances "too extensive and specific to be coincidental." That sentence is doing legal work. It moves the case off the philosophical question everyone likes to argue — does a model "learn" like a person or "copy" like a photocopier — and onto a finding of fact: these outputs reproduce these works, and a model that reliably produces them is a model that contains them.

Two findings, not one

Then the operative reasoning, which has two parts worth separating, because they stand or fall independently. The first is about the model itself. The court held that the disputed songs were, in its phrase, "reproducibly contained" — memorised — inside Suno's system, and that a copy retained in a trained model and stored on a server in Germany is a reproduction within the meaning of Section 16 of the German Copyright Act. Not the output. The model. On this reading, the act of carrying a protected work around inside the weights is already a restricted act, whether or not anyone ever prompts it back out. The second finding is about the output: the generated tracks communicated those works to the public, and the court placed the liability on Suno rather than on the user who typed the prompt. Its reasoning was that "the prompts were simple and open-ended," specifying only lyrics and a desired style, and that it was "the models, not the prompts," that substantively determined what came out. The person asking for "a 1978 disco track" is not the author of the infringement; the system that answers is.

The court did not merely rule that Suno's outputs were too similar to be legal. It ruled that the copies retained inside the model were already the infringement.

That distinction — model as reproduction, separate from output as communication — is more consequential than the result in this one case. It means a claimant no longer has to catch a model in the act of generating a near-identical copy to make out an infringement. If the work can be shown to be recoverable from the system, the reproduction has already occurred, upstream of any prompt. Whether a given model has "retained" its training data in that sense becomes a question of fact, testable model by model — and one that cuts against exactly the systems trained hardest on the most recognisable material.

The defence that was supposed to work

Now the part that will travel furthest, because it concerns the one argument on which every AI developer operating in Europe has quietly been relying. Suno raised the text-and-data-mining exception — Section 44b of the German Copyright Act, which transposes Article 4 of the 2019 EU Copyright Directive. That exception is the legal foundation of the "we can train on anything lawfully accessible" position. It permits reproductions made in order to extract information — patterns, statistics, correlations — from works, unless the rightsholder has expressly reserved the use. It was written for data mining, and the industry has spent five years treating it as a general licence to train.

The Munich court held it did not apply, and the reason it gave is the one that should concentrate minds. Text-and-data mining, the chamber reasoned, permits analysis. It does not permit retention. Once the works were not merely analysed during training but kept — memorised, reproducible — inside the model, the activity fell outside the exception and back under the ordinary reproduction right that requires a licence. Read that slowly, because it is the whole case. The TDM defence assumes training is like reading: the system passes over the works, extracts abstract patterns, retains nothing protectable. The court accepted that framing as the test — and then found, on the technical evidence, that this particular model failed it. The works were still in there.

So Section 44b is not, on this reading, a blanket permission to train. It is a permission to analyse that evaporates the moment a system can be shown to have kept a copy of what it analysed. That reframing hands rightsholders a workable line of attack against models in any medium, not just music: establish, as a matter of fact, that the model retains recoverable fragments of protected works, and the exception the developer was counting on simply falls away. The opt-out mechanics that dominated the last two years of lobbying — who reserved their rights, in what machine-readable form, by when — start to look secondary. If the exception never covered retention in the first place, the opt-out was a fight over the wrong clause.

Training abroad is not a shield

The next move is jurisdictional, and it is the one Suno most likely believed it had insured against. Suno trained in the United States. The conventional wisdom holds that US training is governed by US law, where the fair-use doctrine has so far given AI developers more room than Europe does — and where Suno is separately being sued by the major labels. The Munich court did not defer to any of that. It applied US fair use itself — 17 U.S.C. § 107 — and concluded that fair use did not cover the training here, because the outputs were substantially similar to the originals and were produced from trivial, open-ended prompts. Then, separately, it found the German output infringing under German law. Two infringements, two legal systems, one defendant, one judgment.

What let a German court reach an act committed in America is a provision most coverage will not name: Section 131 of the German Collecting Societies Act, the VGG. It grants German courts international jurisdiction over infringing acts committed abroad — but only when the claimant is a collecting society. That is a narrow and load-bearing door. An individual songwriter, or a single label, could not have pulled US training into a Munich courtroom on this basis. GEMA could, precisely because of what GEMA is. The extraterritorial reach of this judgment is a function of the plaintiff's legal form as much as of the facts — which is worth remembering before anyone generalises the ruling into "European courts can now reach any training run anywhere." A collecting society can. That is not the same sentence.

What the AI Act does not do

There is a second statute the court went out of its way to address, and its treatment corrects a misconception the industry has been cultivating in Brussels for a year. Under the EU AI Act, providers of general-purpose models must adopt a policy to comply with EU copyright law and publish a "sufficiently detailed summary" of the content used for training — the obligations in Article 53(1)(c) and (d). The comfortable reading, in some quarters, has been that discharging these duties buys a kind of safe harbour: disclose the training data, adopt the policy, and the copyright exposure is managed.

The Munich court said the opposite, in terms. Complying with Article 53, it held, "does not in itself release AI providers from copyright liability." It cited Recital 107 of the Act for the proposition that the transparency rules exist to make enforcement easier for rightsholders — not to hand developers a defence. This is the distinction that gets lost every time a compliance obligation is mistaken for permission. The training-data summary is a disclosure that arms the claimant; it is a floor beneath the rightsholder, not a ceiling over the defendant. A model provider that files an immaculate Article 53 summary has told the world, in auditable detail, what it trained on. If some of that was licensed music retained in the weights, the summary is not its shield. It is the claimant's exhibit list.

In force, and now the question of enforced

So what does the judgment actually require, and when does it bite? The court ordered Suno to cease the unauthorised reproduction and training use, to disclose the revenue it earned from the infringement so that damages can be calculated, and to pay those damages once quantified. It is a first-instance decision; Suno has said it disagrees with the findings and is weighing an appeal, and the damages figure is not yet set. There is a genuine open question — flagged in the specialist coverage — about how far the ruling can be enforced while an appeal is pending. German first-instance judgments can often be provisionally enforced against the posting of security, and how aggressively GEMA can press an injunction against Suno's European operations in the interim will matter more, in the near term, than the eventual sum of damages.

But the load-bearing event has already happened, and it is not the injunction or the money. It is the precedent. A European court of record has now held that memorisation is reproduction, that the text-and-data-mining exception does not cover it, that offshore training does not escape it when the plaintiff is a collecting society, and that AI Act compliance does not neutralise it. Any one of those holdings could be narrowed on appeal. Together, on the page, they describe a doctrine — and doctrines, once written down by a court that hears these cases, tend to be cited by the next one.

The licence was always the point

Which brings us to the second-order effect, the one that explains why GEMA fought a technical copyright case with the intensity of a company launching a product. Eight days before the judgment, on 23 July, GEMA opened something called PLAI: a licensing hub offering AI developers pre-cleared access to a catalogue reported at roughly 178,000 tracks across more than sixty genres, with the composition and recording rights bundled together. Read the lawsuit and the licence as a single instrument and the strategy is unmistakable. The litigation establishes that training on European repertoire without permission is infringement. The licence offers that permission for sale. This is the play collecting societies have run for a century — first make the use unambiguously require a licence, then be the only convenient party selling it — now applied to model training. GEMA is not trying to stop AI music. It is trying to be the tollbooth on the road to it.

It helps to see Friday's decision as the second ruling, not the first. In November 2025, the same court found for GEMA against OpenAI, over song lyrics reproduced by ChatGPT. That case was about text; this one is about the music itself, and about the model rather than the output. Two first-instance wins in the same specialised chamber, built on the same theory — that European copyright reaches into the training process and out across borders — start to look less like a pair of results and more like a settled local approach. That is precisely how the Brussels Effect begins: not with a regulation announced to fanfare, but with a court that keeps reaching the same conclusion until everyone plans around it.

No one builds a separate music model for Germany. A developer that must license European repertoire to operate in one of the world's larger music markets will find it simpler to license it everywhere than to maintain a compliant German edition — and simpler still to assume, going forward, that "we trained on the open internet" describes a liability rather than a defence. A civil chamber in Munich cannot bind a court in California or a legislature in Washington, and it has not tried to. But it can change the price of a business model, and it has. The reported version says Germany ruled against an app. The binding version says that anyone who trains a generative model on music the world can hear now has to reckon with a rule written in Munich — and, increasingly, to pay the society that wrote it.