The judgment handed down on 31 July 2026 by the 42nd Civil Chamber of the Munich I Regional Court[1]is one of the most important European decisions to date on generative artificial intelligence and copyright. The decision is highly instructive, both as to the classification of the various stages of the training process and as to their treatment under the conflict-of-laws rules, the exceptions to copyright and the attribution of infringements resulting from generated content.
In January 2025, the German collective management organisation GEMA brought proceedings against Suno, a US company operating an artificial intelligence-based music generator accessible in Germany. It accuses Suno of having used, without authorisation, six protected musical works in the operation of its models, both at the training stage and in certain content generated by the system.
The model was trained in the United States on a corpus of several million recordings, including the six works in dispute. These recordings came from YouTube, from which Suno extracted them using stream extraction techniques (stream-ripping) involving the circumvention of the technical protection measure known as rolling cipher. The audio files were then associated with metadata, including title, genre and lyrics, before the resulting datasets were tokenised and used to train the model. Training consisted of progressively adjusting the model’s parameters and weights.
The dispute concerns less the factual reality of these operations than their legal classification, and chiefly whether the works remained embedded in the model in a form amounting to a reproduction. Having no access to the model’s parameters, GEMA relies in particular on a comparison between the original works and the generated content. Its representatives entered the lyrics, the title and a genre indication into the interface, without any instruction as to melody, harmony, rhythm or arrangement, and obtained tracks in which the compositions remained recognisable. GEMA argues that the musical similarities observed show that the elements reproduced come from the model itself and not merely from the instructions provided by the user. The lyrics, however, are not the subject of the action. Only the musical works are claimed.
The Court largely upholds GEMA’s claims. It finds separate infringements relating, on the one hand, to the reproductions made in the United States for training purposes and, on the other, to the memorisation of the works in the model operated in Germany and to the reproductions and communications to the public resulting from the operation of the system. The operative part prohibits Suno from reproducing the works in dispute for training purposes in the United States, from storing them in Germany “in the form of parameters or other data structures” in its models, and from generating content reproducing the protected elements of the works claimed. The Court further orders Suno to provide the information needed to determine the extent of the infringements and declares that it is liable to compensate GEMA, the amount of damages remaining to be determined.
The common thread of the judgment lies in breaking down the generative process into several legally distinct operations. The Court distinguishes, first, the reproductions made in the United States to build the training data and train the model, governed by the US Copyright Act and the fair use analysis; second, the memorisation of the works in the model’s parameters, treated as a separate reproduction; and finally, the reproductions that may result, in Germany, from generated content where it reproduces protected elements of the original works. The generative process is therefore not treated as a single act, but as a succession of operations that may each be classified in their own right and, where appropriate, be governed by different laws.
This method enables the Court to separate the acts of reproduction territorially, to treat memorisation in the model as a separate act, and then to assess separately the justifications that may cover the various operations, whether the text and data mining exception, US fair use or arguments drawn from the Artificial Intelligence Act (AI Act).
We will therefore examine, first, the plurality of acts of reproduction in the generative process (I), before considering, second, the effectiveness of the exclusive right as regards the generative model (II).
I. The plurality of acts of reproduction in the generative process
The Munich decision rests on breaking down the generative process into several legally distinct acts. This approach leads the Court, first, to separate the reproduction operations territorially according to where they took place (A), and then to recognise the reproduction resulting from the memorisation of the works in the model as a separate act (B).
A. The territorial separation of acts of reproduction
The fragmentation of the generative process manifests itself, first, in relation to jurisdiction and the applicable law. Suno is established in the United States, where the models in dispute were trained, while those same models are operated in Germany, in particular by means of servers located on German territory. The dispute therefore raises questions both as to the German court’s jurisdiction to hear acts carried out in the United States and as to the law applicable to the various reproductions relied on.
The Munich Regional Court first finds that it has jurisdiction to hear all of GEMA’s claims, including those relating to the reproductions made in the United States for training purposes. For infringements committed in Germany, its jurisdiction derives from Section 131(1) of the German Collective Management Organisations Act (VGG) and Section 32 of the German Code of Civil Procedure (ZPO). GEMA argues that the generator in dispute is accessible throughout Germany and, in particular, in Munich, where the alleged infringements can be located. The question is more delicate for the acts of reproduction carried out in the United States. To include them in the same dispute, the Chamber applies the concentration rule laid down in Section 131(2) VGG. This allows a collective management organisation, where infringements attributable to the same infringer fall within the jurisdiction of different courts, to consolidate its claims before one of the competent courts.
Under the principle of double functionality[2], the Court gives this rule international scope by holding that the German court with jurisdiction over infringements committed in Germany may also rule on acts carried out abroad where they form part of the same set of disputed facts. The solution is remarkable, but its scope should be kept in perspective. The mechanism rests on a provision specific to collective management organisations, intended to facilitate the collective exercise of rights by avoiding a multiplicity of proceedings. The Court itself emphasises the particular status of these organisations to justify this concentration of jurisdiction. The decision therefore does not establish, as a general rule, universal jurisdiction of the German courts over any training activity carried out abroad.
The jurisdiction of the German courts does not entail any corresponding extension of German law. The Chamber is careful to keep the two questions clearly separate. It rejects Suno’s argument that the principle of the country of protection precluded the German court’s jurisdiction. The lex loci protectionis does not determine which court has jurisdiction, but the law applicable to the merits. This solution has its source in the principle of territoriality of intellectual property rights, under which protection is assessed under the law of the country for which it is claimed. The German court can therefore hear acts carried out in the United States without subjecting them to German law.
This distinction also applies to GEMA’s standing to sue. Ownership resulting from a contractual assignment is governed not by the lex loci protectionis, but by the law applicable to the contract. Applying Article 4(2) of the Rome I Regulation, the Court holds German law applicable, the characteristic performance being that provided by the collective management organisation. The chain of title can thus be assessed under German law, including for acts located abroad.
The Court then applies a distributive approach to the law applicable to the acts complained of. The reproductions made in the United States for training purposes are governed by US law, while the acts of reproduction carried out in Germany are governed by German law. The Chamber refuses to treat the generative process as a single operation governed by a single law. Each act retains its autonomy and must be connected to the territory in which it is carried out.
This solution matches one of the alternatives identified in the report submitted to the French Higher Council for Literary and Artistic Property (CSPLA) under the chairmanship of Professor Tristan Azzi[3]. The report contrasts an approach based on a plurality of simple torts, leading to a distributive application of the laws of the various places of reproduction, with an approach based on the existence of a complex tort, making it possible to treat a chain of legally interdependent events as a whole.
The Munich Court adopts the first approach. This solution nonetheless raises particular difficulties in the field of generative artificial intelligence. Locating training activities on the basis of technical infrastructure or server location may prove uncertain and impractical, since the various operations needed to feed a model may be spread across several territories[4]. Choosing such a criterion also allows the model provider to influence indirectly the determination of the applicable law by choosing to locate its servers in a State offering a lower level of copyright protection. The CSPLA report therefore favours a unitary approach, based on the indivisibility of the process linking input and output, which would locate the exploitation at the stage of the damage, in the country where the model is marketed and where users access the generated results.
The divergence between these two approaches is particularly visible in the Suno case. Applied to the facts of the case, the unitary approach could have led to the upstream reproductions made in the United States also being governed by the law of the country where the model is marketed.
B. Memorisation as a separate act of reproduction
The fragmentation of the generative process continues on the substantive level. The Court distinguishes the reproductions needed to build the corpus, the analysis and training of the model, and then its use by means of prompts and generated content. Above all, it contrasts the reproductions made to enable the analysis of the data with those that remain in the model after training. Where elements of the work remain embedded in its parameters in a reproducible form, their memorisation may constitute a separate act of reproduction.
The solution is not entirely new. A few months earlier, in GEMA v. OpenAI of 11 November 2025, the Munich Regional Court had already adopted similar reasoning regarding the memorisation of works in an artificial intelligence model. The Suno judgment builds on that precedent, but transposes its reasoning from textual content to musical works. This continuity tends to confirm that, for the Munich judges, memorisation must be considered independently of the copies made upstream to build the corpus and train the model.
Suno specifically disputes that such a reproduction can be established. In its view, the model’s parameters do not contain the works themselves, but only statistical relationships and general characteristics learned during training. The model’s capacity would, moreover, be insufficient to physically store all the data that made up the corpus.
The Court rejects this argument. Classification as a reproduction does not require the work to be kept in the form of an identifiable file, nor that it can be reproduced identically. It is sufficient that the parameters incorporate more than mere statistical regularities or correlations and allow recognisable elements of the work to be reproduced.
The reasoning rests on the broad conception of the reproduction right derived from Article 2 of Directive 2001/29, implemented in German law by Section 16 of the German Copyright Act (UrhG). The Chamber points out, referring in particular to the Infopaq judgment[5], that the concept of reproduction must be given an autonomous and uniform interpretation in EU law. Since Article 2 covers reproduction “by any means and in any form”, the technique used to fix the work is irrelevant. In the name of technological neutrality, the Court likens memorisation in the parameters to the digital storage of a musical work in compressed form, such as an MP3 file. Changing its digital representation is not enough to rule out reproduction.
Furthermore, the Chamber holds that fixation may be established where the work is not directly accessible in the model but remains indirectly perceptible thanks to the information incorporated in it. The dispersal of this information across the various parameters is therefore no obstacle, provided that the model, taken as a whole, makes it possible to reproduce recognisable elements of the work. Reproduction results not from the retention of an original file, but from the system’s ability to make the work perceptible through its operation.
This classification nonetheless remains subject to proof of actual memorisation of the works in dispute. In this case, that proof is drawn in particular from a comparison between the works and the results obtained from simple, open prompts. The correspondences found in melody, rhythm or harmony are held to be too significant to be explained by chance, whereas the prompts did not contain sufficiently precise musical indications to determine those elements themselves.
II. The effectiveness of the exclusive right as regards the generative model
Having identified the various acts that may fall within the exclusive right, the Court examines the conditions under which they can be enforced against the model provider. It first sets the reproductions made for training against the mechanisms that might justify their use (A), before analysing the attribution of infringements resulting from the system’s generation of content (B).
A. The exclusive right maintained at the training stage
Classifying the various reproductions is not enough to establish that they are unlawful. It must also be determined whether they can benefit from the mechanisms that may justify the use of protected works for training purposes. Here again, the separation made by the Court drives the reasoning. Reproductions governed by German law are tested against the text and data mining exception, while those made in the United States are assessed under fair use. In both cases, the Chamber refuses to infer from them a general authorisation to train a model on protected works.
As regards text and data mining, the Suno judgment follows directly from GEMA v. OpenAI. The Court does not, as a matter of principle, refuse to apply Article 4 of Directive 2019/790 (DSM) and Section 44b UrhG to artificial intelligence models. On the contrary, it accepts that reproductions intended to prepare the training corpus may fall within text and data mining. But the exception remains tied to its purpose, namely enabling the automated analysis of data in order to extract information, patterns or correlations from it. Where elements of the work are, beyond that analysis, embedded in the model’s parameters in a reproducible form, memorisation constitutes an additional reproduction that is no longer covered by the exception.
In addition to this first ground, the Suno case raises a difficulty relating to lawful access to the works. The Chamber finds that the company obtained the files from YouTube by stream-ripping, circumventing the rolling cipher mechanism designed to prevent their download. Yet Article 4 of the DSM Directive and Section 44b UrhG presuppose lawful access to the works. The Court therefore holds that this condition was also not met.
Nor does the Artificial Intelligence Act make it possible to neutralise this conclusion. Suno argues that the obligations imposed on providers of general-purpose AI models by Article 53(1), in particular the adoption of a policy to comply with copyright and the publication of a sufficiently detailed summary of the content used for training, necessarily implied the possibility of marketing such models without obtaining a licence for each of the works used. The Court rejects this interpretation. The AI Act refers back to existing copyright law, and its recital 108 states that it does not affect its application. As for the transparency obligations in Article 53, they are intended precisely to enable rightholders to identify the content used and to exercise their rights. They constitute neither a new exception nor a regime authorising training.
The examination of US law leads the Chamber to a comparable conclusion. The reproductions made in the United States are assessed under fair use as provided for in § 107 of the Copyright Act. The Court does not rule out that training a generative model may, in principle, be “transformative” in a way that may fall within fair use, but it stresses that this classification depends on the specific circumstances of the use. In this case, the memorisation of the works and the generation of content substantially similar to them lead the Court to qualify that transformative purpose. Added to this are the commercial nature of the exploitation and the conditions under which the files were acquired by stream-ripping.
These circumstances allow the Chamber to distinguish the US decisions Bartz v. Anthropic and Kadrey v. Meta, in which no output substantially similar to the training works had been established. They also weigh in the assessment of the effect on the market. The generated content, which can be listened to free of charge, stored and distributed, may substitute for the original works on the market and affect their economic value. Weighing up the factors leads the Court to reject fair use, without however laying down the principle that training artificial intelligence models is generally unlawful.
B. Attribution of infringements at the generation stage
The fragmentation carried out by the Court finds its final extension in the generated content. After the reproductions needed for training and the one resulting from memorisation in the model, the outputs may in turn constitute acts of exploitation where they reproduce protected elements of the original works. A difficulty specific to generative systems then arises: is the user’s intervention by means of a prompt sufficient to attribute the infringement to the user, or can it be attributed directly to the model provider?
Suno argues that the disputed results are attributable to GEMA’s representatives, who deliberately prompted the generator in order to obtain tracks close to the works claimed. The Court rejects this analysis. The prompts did contain the lyrics, the title and general indications of style, but no instruction as to melody, harmony, rhythm or arrangement. Repeating them, sometimes up to one hundred and seventy-six times, did not amount to progressively steering the system either. The Chamber therefore considers that the disputed musical features derive mainly from the model, whose training data Suno compiled, whose architecture it determined and whose training it carried out. The use of a simple, open prompt is therefore not sufficient, in these circumstances, to break the chain of attribution to the provider.
This analysis leads, first, to a finding of acts of reproduction at the generation stage. The results are fixed in the memory of the devices and on the servers used to produce them, within the meaning of Section 16 UrhG. The temporary reproduction exception in Section 44a UrhG, transposing Article 5(1) of Directive 2001/29 InfoSoc, does not apply here. The storage has independent economic significance, since it allows in particular the track to be listened to while it is being generated and contributes to the operation and attractiveness of the service. The conditions for the exception are therefore not met.
The Chamber then finds a communication to the public, but on the basis of a particular construction. It first rejects the specific classification of making available to the public under Section 19a UrhG, since GEMA had not sufficiently established that the works could be accessed at a time chosen by the user. It nevertheless upholds the alternative claim based on the unnamed right of communication to the public under Section 15(2) UrhG, interpreted in accordance with Article 3(1) of the InfoSoc Directive. Offering the model, using it to generate music and the disputed outputs may constitute acts of communication to the public. The public nature stems in particular from the service being open to an indeterminate and sufficiently large number of users.
French law provides a more direct basis on this point. Article L. 122-2 of the Intellectual Property Code broadly defines performance (représentation) as the “communication of the work to the public by any process whatsoever”. Unlike German law, this definition is not based on a list of modes of exploitation, but covers in general terms any communication of the work to the public, whatever the technical means used “to establish the relationship between the public and the work or protected subject matter”[6]. The generation and transmission of content reproducing protected elements may fall directly within the performance right, without any need to resort to a residual exploitation right.
Finally, Suno cannot benefit from the hosting exemption regime provided for in Article 6 of the Digital Services Act (DSA). The company does not merely passively store information provided by its users. The model and the application it operates contribute directly to the production of the disputed content. The judgment links the infringements found at the generation stage to the very operation of the system. Where the user does not determine the protected elements reproduced and they result mainly from the model’s architecture, training and memorisation, their generation can be attributed directly to its provider.
The Munich judgment thus adopts a particularly detailed approach to applying the exclusive right to generative models, distinguishing the various operations involved in their training and exploitation.
This decision comes against a backdrop of parallel development of contractual arrangements between rightholders and artificial intelligence model providers. Less than two weeks after the judgment, Suno announced the conclusion of a licensing agreement with BMG, an international music publishing and rights management group. A few days before the decision, GEMA had launched “PLAI”, a platform offering licensed, structured and scalable datasets specifically intended for training AI systems. These initiatives reflect a growing trend towards contractual arrangements for access to the works used to train generative models.
Beyond its legal contribution, the Munich decision helps open up a discussion on defining the conditions under which rightholders can authorise, regulate and derive value from the use of their works by generative models.
[1] Case No. 42 O 763/25
[2] In German law, the theory of double functionality (Doppelfunktionalität der Zuständigkeitsnormen) means that domestic rules on territorial jurisdiction serve at the same time as rules on international jurisdiction.
[3] T. AZZI (chair), Rapport de mission sur la loi applicable, en vertu des règles de droit international privé, aux modèles d’intelligence artificielle générative commercialisés dans l’Union européenne, CSPLA, 15 Dec. 2025.
[4] CSPLA report, cited above, No. 32 et seq.
[5] ECJ 16 July 2009, C-5/08, Infopaq, para. 41 et seq.
[6] A. Lucas and A. Lucas-Schloetter, Traité de la propriété littéraire et artistique, LexisNexis, 6th ed., 2026, No. 724, p. 680.
