GEMA v Suno: what the Munich judgment really means for AI-assisted music

The Munich Regional Court did not declare AI-assisted music unlawful. It did, however, place one question at the centre of the debate: who controls the material used, the outputs produced and the chain of rights?

The judgment delivered on 31 July 2026 by Munich Regional Court I, the Landgericht München I, in GEMA v Suno, case 42 O 763/25, requires a practical rather than ideological analysis. The issue is not whether artificial intelligence is inherently good or bad. It is which works were used, on what legal basis, what remained reproducible within the model, what appeared in the generated results and who was responsible for them.

The specialist 42nd Civil Chamber largely upheld GEMA’s claims for an injunction, disclosure and damages.

The proceedings concerned six musical works:

One distinction is essential: alleged infringement of the lyrics was not part of these proceedings. The dispute concerned the musical works and compositional elements that the court considered recognisable in the generated material.

According to the court’s official statement, it was undisputed that the six works had been included in the training material. The recordings had been extracted from YouTube using stream-ripping techniques that circumvented the platform’s “Rolling Cipher”, a technical measure intended to prevent audio and video downloads.

The court found that the works were reproducibly contained in versions v3.5 and v4 of the model, which were stored on servers in Germany.

Under the reasoning summarised in the official release, memorisation occurs where a model does more than derive general information from training data. Content from the works remains incorporated in the post-training parameters in a form capable of re-emerging through outputs.

The court compared the works contained in the training data with the generated results. In view of the length and complexity of the compositions, it rejected coincidence as the explanation for the similarities. The memorisation was therefore treated as an infringement of the reproduction right under section 16 of the German Copyright Act, the UrhG.

To generate the disputed outputs, GEMA entered the title of each work, its original lyrics and the requested musical style. The prompts did not specify melody, harmony, rhythm or arrangement.

The court described those prompts as simple and open-ended. It therefore held that the user’s conduct did not break the causal link between the model and the resulting content. Responsibility was attributed to the party that selected the training material, operated the model, controlled its architecture and was responsible for memorisation.

This does not mean that users can never be liable for what they request, select or publish. It means something narrower: a provider may not always shift responsibility to users where recognisable protected music emerges without detailed musical instructions being supplied through the prompt.

The official statement contains a further significant finding. It says that, in the circumstances considered, the offer of the model and application for generating music itself infringed an unnamed form of communication to the public under section 15(2) UrhG.

The release does not contain the court’s complete reasoning. That finding should therefore not be converted into a general rule governing every generative music model.

Article 4 of Directive (EU) 2019/790 provides, subject to conditions, an exception or limitation for reproductions and extractions of lawfully accessible works made for the purposes of text and data mining. The exception also depends on the use not having been expressly reserved by the rightholder in an appropriate manner.

In GEMA v Suno, the court held that the reproducible presence of the works in the model was not covered by section 44b UrhG, the German text and data mining provision. The distinction is fundamental: analysing a work to extract information is not necessarily the same as retaining expressive material in a form capable of being reproduced through outputs.

The proceedings also concerned copies made during training in the United States. The court asserted jurisdiction under section 131 of the German Collective Management Organisations Act, the VGG, applied US law and rejected the fair-use defence. It distinguished the case from decisions in which training material had not been made substantially available to users through model outputs. Here, according to the release, simple and open-ended prompts produced outputs substantially similar to the original works, and all four fair-use factors weighed against the defendant.

This is not a universal finding that every form of AI training falls outside US fair use. It is an assessment of the evidence, works, model versions and outputs considered in this litigation.

The court expressly stated that the judgment is not final. GEMA’s claims were upheld “largely”, but the official statement does not identify every claim that was rejected, limited or only partly allowed.

The publicly available official source is the court’s press release rather than the complete reasoned judgment. It would therefore be unsafe to state as fact the full wording of each injunction, every territorial limitation, the final calculation of damages, the terms of any provisional enforcement or the precise claims that were unsuccessful.

Suno told Reuters that it was considering all available options, including an appeal. As at 5 August 2026, the official sources reviewed do not confirm that an appeal has been filed.

Some news coverage has described the decision as “historic” or “landmark”. Those are editorial assessments. The legally more precise description is a significant first-instance judgment that may influence future disputes but remains open to appellate review.

Article 53 of the EU AI Act requires providers of general-purpose AI models to adopt a policy for compliance with Union copyright law, respect rights reservations expressed under Article 4 of Directive 2019/790 and publish a sufficiently detailed summary of the content used for training.

The Munich court did not list the AI Act as a legal basis for this judgment. It is relevant regulatory context, not the rule on which the court states that the decision was founded. Its Article 53 obligations also do not automatically apply in the same way to every professional user or downstream operator.

The court did not prohibit AI-assisted music or make every musical similarity unlawful: copyright does not generally monopolise a genre, atmosphere or broad musical idea. The judgment does not clear every layer of the rights chain, nor does it remove the responsibility of those who select, approve, publish or use content professionally.

Asking whether “AI music is legal” usually produces an answer that is too broad to be useful. A professional review should distinguish at least three layers:

A downstream licence does not automatically cure an upstream infringement. Equally, an original recording cannot be exploited safely unless the licensor controls the rights required for the intended use.

Direct licensing and collective management are different methods of administering music rights. A direct licence can be effective where the specific chain of authorisation permits the grant, taking account of the right concerned, territory, intended use and any mandates granted. Collective management organisations perform an equally important function where rights have been entrusted to collective administration.

The compliance question is not “direct licensing or a collecting society?”. It is: who controls these particular rights for this use and territory, and can that control be demonstrated?

Within the MoosBox model, coverage depends on the documented chain of authorisation applicable to the content and on the contractual terms of the service. MoosBox is not a Collective Management Organisation or an Independent Management Entity. It operates as a music service provider on the basis of direct licences, agreements with duly entitled parties and content for which it holds the authority required for the authorised uses.

The MoosBox music licence applies to the Client, authorised locations, territories, uses and period, and remains valid while the subscription is active and payments are up to date. The Licence Certificate is issued after the first payment and final activation; the trial is technical and does not include a commercial licence for public performance. Coverage applies to content supplied through the service, not music from Spotify, YouTube, radio, mainstream services or other external sources.

For us, technology is a creative instrument. It is not an editorial shortcut.

The MoosBox approach to AI-assisted music is based on human direction, selection, listening, traceability and a genuine ability to reject content. Briefs are not designed to reproduce a particular song, recognisable melody, identifiable voice or named artistic identity.

Imitation prevention begins before generation, in the way the brief itself is written.

Before inclusion in the service, AI-assisted content is subject to documented selection, review, reasonable similarity checks against known or recognisable works, and traceability procedures.

These procedures are intended to prevent and manage risk. They do not constitute a certificate of originality or an absolute guarantee that similarities or third-party claims cannot arise.

Content presenting intentional references, unresolved doubts or similarities that have not been adequately clarified is not approved for production or distribution.

Responsible governance should make it possible to reconstruct the main stages of the process: the version reviewed, checks undertaken, outcome and editorial decision. Traceability does not eliminate every risk, but it prevents content from entering a catalogue through a blind process for which nobody is accountable.

Where a substantiated claim or reasonable doubt arises concerning specific content, MoosBox may suspend, remove or replace that content as a precaution while checks are carried out. Subject to the contractual terms, this does not affect the validity of the licence for the remainder of the catalogue.

The wider framework is set out in the 2026 MoosBox Legal Protocol and our guide to AI music ownership, licensing and compliance.

No. It decided a specific dispute involving six works, two model versions and identified outputs. It did not impose a general prohibition.

No. The court expressly stated that it is not final. The defendant has said that it is considering an appeal.

Under the court’s reasoning, the model did not merely learn general characteristics. Content from a work remained incorporated in a sufficiently reproducible form to emerge through outputs.

No. Article 4 of Directive 2019/790 imposes conditions, including lawful access and the absence of an effective rights reservation. The court also held that the reproducible memorisation at issue was not covered by section 44b UrhG.

No. Checks do not certify originality; a direct licence covers only rights genuinely controlled by the licensor and cannot automatically cure an infringement occurring earlier in the process.

The most useful dividing line is not between human music and AI-assisted music. It is between governed processes and opaque ones.

On one side are systems in which nobody can clearly explain what material was used, what remained reproducible or why an output was released. On the other are organisations that direct, listen, review, document, reject and accept responsibility for their decisions.

Technology may help us create more quickly. Trust still requires method, transparency and people willing to be accountable for the choices they make.

Updated 5 August 2026. General information only and not a substitute for legal advice on a particular matter, territory or repertoire.

Sources reviewed and verified on 5 August 2026