On 31 July 2026 the Regional Court of Munich I decided the first European case brought against a generative music tool, ruling largely for the German collecting society GEMA against Suno Inc. The court prohibited four acts in relation to six musical compositions: reproduction for training purposes in the United States, reproduction through memorisation in the model in Germany, communication to the public through the offering of the model, and reproduction and communication to the public through the outputs. It ordered Suno to disclose the scale of its use and found the company liable in damages. The judgment is a first-instance decision and it is not final. Coverage recorded it as a European court holding that training a music model on copyrighted work is unlawful, which is roughly half of what happened. Two of the four prohibited acts concern the model itself rather than the music it produces, and both of those turn on memorisation, a word carrying no agreed test in copyright law or in machine learning.
What the court actually prohibited
The four acts separate cleanly into three stages of the same pipeline. One is the copying that happens when works are ingested for training. One is the presence of those works inside the trained model, which the court treated as an ongoing reproduction rather than a spent act. The last two cover the product built on that model: offering it to the public in Germany, and the outputs users generate with it.
Six compositions carried the case, which is a small number for a judgment being read as a verdict on an industry. The court held that the burden of disproving market harm sat with Suno, declined to make GEMA produce revenue figures, and found that substantially similar outputs available free on the entry tier substitute for the originals. That reasoning matters more than the headline, because it sets who has to prove what the next time a rightsholder sues.
A German court applying American law
The training limb is the strangest part of the decision and the least discussed. Suno trained in the United States, so a German court had to reach an act that happened abroad. It took jurisdiction under a venue rule available to collecting societies, applied United States copyright law to those acts, and rejected fair use on the facts before it.
A German judgment on American fair use is unusual enough that it is expected to be the centre of any appeal, alongside the memorisation standard and the 176 prompts GEMA used to produce the outputs it complained of. None of that sits comfortably in a headline about training being illegal in Europe, because the European part of the ruling is the memorisation limb, and the training limb was decided under the law of another country.
Memorisation is carrying the weight
The Munich court has now used this reasoning twice. On 11 November 2025 the same court held that song lyrics memorised inside OpenAI models infringed, in a case brought by GEMA over German works including Atemlos and Männer. It rejected the argument that training data is not stored in a model, noting that research shows training data can sit inside a model and be extracted through its outputs, and it held that the text and data mining exception in section 44b of the German Copyright Act does not cover training of this kind.
Missing from both rulings is a threshold. Nobody has said how much of a work has to be recoverable, under how much prompting, before a model counts as containing it. Suno was tested with 176 prompts. A standard whose answer moves with the number and design of the prompts is a standard about the examination rather than about the thing examined, which is why that number is expected to be attacked on appeal. Model weights are also not a filing cabinet, and the informal picture of a song sitting in storage inside a model does not describe how these systems represent anything. The law now needs a test for a property that engineering measures only by trying to extract it.
Platforms are policing an entirely different question
Every gate the music industry has built for this problem so far asks how a track was made. Beatport screens fully AI-generated tracks at upload and keeps the AI-assisted ones, which puts the weight on the undefined threshold of majority human-made. ARIA made wholly AI-generated recordings ineligible for its charts and relies on what artists declare. Spotify badges generated identities and marks who an artist is rather than how the music was produced. All three are answered by a person filling in a form at upload.
Munich asked something no upload form can ask. Whether a model contains protected works is a fact about the model, discoverable only by testing it or by making its operator disclose, which is why the court ordered disclosure of the scale of use. The two regimes can come apart in both directions. A track declared fully human could come out of a model a court finds infringing, and a track flagged as AI-assisted could come from a model trained entirely on licensed material. Labelling schemes describe the track. The Munich cases describe the tool, and the tool is where the money and the liability sit.
What is still open
The text and data mining question that both Munich rulings answer for themselves is pending above them. The Federal Court of Justice has the LAION case, and the Court of Justice of the European Union has Like Company v Google, with an Advocate General opinion due on 3 September 2026. The Munich court saw no need to refer its own case, so its interpretation of section 44b stands only until one of those produces a different one.
For anyone booking rooms and selling tickets, none of this changes what happens at a door this weekend. It changes what a catalogue is worth and who has to license what, and those are the terms on which the next decade of recorded music gets priced. Watch the appeal for whether memorisation survives as a legal standard, and watch 3 September for the first sign of what the European court makes of the exception underneath all of it.