GEMA v. Suno: Training an AI Model Without a Licence Is Reproduction

On 31 July 2026 the 42nd Civil Chamber of the Landgericht München I (Munich I Regional Court) prohibited the US provider Suno from reproducing protected musical works — expressly including their use as training material. GEMA, the German collecting society for music rights, largely prevailed: injunction, disclosure, damages. The works at issue included titles such as “Forever Young” and “Atemlos”.

It was reported as a win for the music industry. The reasoning is more interesting: storing protected works inside the model is an act of reproduction under § 16 UrhG (the reproduction right in the German Copyright Act), and the text-and-data-mining exception in § 44b UrhG does not cover it. That concerns every company using generative AI, and it raises two questions procurement processes never ask: what was my tool trained on — and who owns what it produces?

At a glance

• What was decided: Landgericht München I, 42nd Civil Chamber, 31 July 2026 — training without a licence is reproduction (§ 16 UrhG), and § 44b UrhG does not apply.
• What follows: European courts have jurisdiction even where training happened in the US, as long as the system is operated in Europe. The licensing risk travels down the supply chain, all the way to whoever uses the output.
• The catch: Not final, and enforceable only against security; the parallel case GEMA v. OpenAI is on appeal before the Oberlandesgericht München (Munich Higher Regional Court).

What the court decided

Memorisation is the core. GEMA did not merely assert that works had gone into the training; it showed that the generator produced songs closely resembling the originals. From that the chamber inferred that the works are stored inside the model, and it saw an infringement of the reproduction right in that storage on servers in Germany. What is attacked is not the analysis, but what remains in the model afterwards.

That also disposed of § 44b UrhG. The provision permits text and data mining and the copies needed for it — copies that must be deleted once the analysis no longer requires them. A model that permanently stores works in recognisable form has left the analysis behind; that, in the chamber’s view, is where the exception ends. It is the second Munich judgment of its kind:

	GEMA v. Suno	GEMA v. OpenAI
Court	LG München I, 42nd Civil Chamber	LG München I; appeal: OLG München
Date	31 July 2026	11 November 2025, appeal filed 8 December 2025
Subject	musical works	nine song lyrics
Core	memorisation is reproduction, § 44b UrhG does not apply	infringement in training and operation
Status	not final	not final, pending (OLG München, 6 U 3662/25 e)

Two chambers, two defendants, the same direction. That is the solid statement — and expressly not the statement that the law is now settled.

Three reasons not to overstretch the judgment

First: one chamber, one instance, one case. The Oberlandesgericht München may decide the appeal differently, and neither the Bundesgerichtshof (Federal Court of Justice) nor the Court of Justice of the EU has ruled on how far § 44b UrhG reaches in model training. Anyone telling you today how this ends is speculating — even if they speculate in the same direction as two Munich chambers.

Second: the case was carried by the output, not by the training. Use as training material was expressly prohibited, but the evidence came from the other side of the model — from outputs that audibly resembled the originals. In practice that means: your most reliable indicator is not the training-data list, but what the tool actually produces for you. That disclosure was awarded as well shows that the courts see how hard it is for rightsholders to prove their case.

Third: music is the most sensitive case, not the typical one. A music repertoire is tightly organised in legal terms — a collecting society bundles the rights, and similarity is audible. With text, images and above all code, the chain of rights is more diffuse and similarity harder to pin down. That does not change the law, but it does change what can be proven.

The same shape, passed along three times: from the training material into the model, into the output, into your files. The mark sits on the last repetition — where the licensing risk arrives.

What this means for your company

Ask about provenance — and document the answer. You will not get a complete training-data list, and you do not need one. What is useful is what the provider commits to in writing: that training data is licensed, that machine-readable reservations of use are respected, that outputs do not fall back on memorised works. A “we don’t comment on that” is also a result and belongs in the risk register — which never lists shadow AI, because nobody knows it is there.

Negotiate indemnity, not assurances. The clause that matters most is not a representation about training data but the indemnity: who bears defence costs and damages if a third party brings claims over an output? Check caps, carve-outs and control of the proceedings — an indemnity capped at the annual fee will not hold. The five points in context: AI contracts, what really belongs in them.

Secure your outputs in both directions. Outward: a review wherever recognisability is a realistic risk — music, image motifs, longer passages of text, code from well-known repositories. Inward: pure AI output is usually not protected by copyright (§ 2 (2) UrhG); anyone wanting exclusive use of results needs contractual rights of use and should document their own creative contribution. The AI Act does not help here: its deadlines concern transparency and risk classes, not rights in the training material.

Conclusion

The judgment is not a turning point you can build a strategy on — the instance is too low and the case too open. Nor is it an industry curiosity you can wait out. What shifts is the negotiating position: as long as § 44b UrhG counted as a free pass for commercial training, asking about training provenance was futile. Now it is legitimate.

For mid-sized companies this calls for little panic and one clear task: you cannot influence the law on training, but you can influence your tool inventory, your contracts and your output review. Whether your contracts hold up under this, let’s talk it through. I read them as a business lawyer and build the systems in question myself.

FAQ

Is training AI on protected works now prohibited in Germany?

Not in such sweeping terms. One chamber of one regional court decided one case, and the judgment is not final: training without a licence is reproduction under § 16 UrhG, and § 44b UrhG does not cover it. The parallel case against OpenAI is on appeal before the Oberlandesgericht München — what is solid is a direction, not a settled legal position.

Are we liable if we use an AI tool whose training data is unclear?

The injunction binds the provider first, not the user. The risk travels on, though: whoever uses an output that recognisably reproduces a protected work is reproducing that work in their own business. More important than the training-data list you will never receive are therefore a robust indemnity clause and a review of outputs wherever recognisability is a realistic risk.

What does § 44b UrhG cover, and why did it not hold here?

§ 44b UrhG permits text and data mining, meaning the automated analysis of digital works; the copies needed for it must be deleted once the analysis no longer requires them. Rightsholders may reserve their works from such use, machine-readably in the case of works available online. What carried the case was that the model had memorised works and could output them recognisably — so the reproduction did not end with the analysis.

Who owns what our AI tool produces?

Pure AI output is usually not protected by copyright in Germany, because § 2 (2) UrhG requires a personal intellectual creation. Contract for rights of use rather than an assignment of copyright that has nothing to assign. Conversely, the absence of your own protection does not mean that no third-party rights attach to the output — both questions belong in the same contract.

---

Sources — as of 12/08/2026

• GEMA on the Suno decision — https://www.gema.de/de/w/suno-entscheidung-2026
• LTO on the Landgericht München I judgment — https://www.lto.de/recht/hintergruende/h/landgericht-muenchen-i-gema-suno-ki-modell-training-songs-musikkomposition-urheberrechte
• Bird & Bird on the decision — https://www.twobirds.com/en/insights/2026/germany/munich-district-court-rules-on-ai-generated-music-gema-v-suno
• CMS on GEMA v. OpenAI — https://cms.law/de/deu/legal-updates/gema-vs.-openai-urheberrechtliche-grundsatzentscheidung-des-landgerichts-muenchen-i-ergangen
• urheber.info on the appeal filed — https://urheber.info/diskurs/openai-legt-berufung-ein

On the sources: GEMA’s announcement confirms the chamber and the date of the judgment but names no case number, so case numbers resting only on secondary sources are omitted here.

This article is general information and not legal advice in an individual case. As of 12 August 2026; both proceedings are open, so please check the current state before making decisions.