Creation of music with generative AI may be copyright infringement – Munich Regional Court rules in GEMA’s lawsuit against Suno

Key Takeaways

• Memorisation constitutes reproduction: If training works are recognisable in the outputs, the court considers this sufficient evidence that they are stored in the model and are being reproduced pursuant to Article 2 of the InfoSoc Directive.
• TDM exception does not cover memorisation: Article 4 of the DSM Directive permits preparatory reproductions for building the training corpus, but does not extend to reproductions that enable the subsequent reproduction of the works.
• EU AI Act does not create a copyright safe harbour: Compliance with the transparency obligations under Article 53 of the EU AI Act does not replace the need for licensing
• No fair use under 17 U.S.C. § 107: The training of the AI violated U.S. copyright law because the works used were recognisable in the output.

In a highly anticipated decision, the Regional Court of Munich I (the ‘Court’) recently ruled in a case against the provider of the AI music generator Suno, determining under what conditions the generation of music based on existing songs constitutes copyright infringement (Regional Court of Munich I, judgment of 31 July 2026, Case No. 42 O 763/25). The decision is particularly notable because the court distinguishes between reproductions made during training in the United States, reproduction through memorisation in model parameters in Germany, and infringements resulting from the model’s deployment and outputs in Germany.

The Court found that the songs used for training were still memorised in the model itself and could be reproduced after the entry of relatively generic prompts. The Court regarded the recognisability of the works in the generated outputs as evidence that the related training data was embodied in the model and not merely information extracted during training that is not protected by copyright. This was deemed to constitute a reproduction of the works. The provision of generated songs that showed sufficient similarity to the songs used for training was also deemed to constitute an unlawful communication to the public.

Furthermore, the Court dismissed the defence based on the exception for text and data mining. While preparatory reproductions for analytical purposes may generally be exempt, this would not apply to the permanent storage of copyrighted works within the model. In addition, the Court found that the training data had been obtained by circumventing the technological protection measures of YouTube.

Finally, the Court also examined the legality of the acts of reproduction carried out in the United States in connection with the training of the models, applying U.S. copyright law. The Court treated these acts as infringements of the U.S. right of reproduction under 17 U.S.C. § 106(1). It ruled that the use was not justified by the fair use doctrine (17 U.S.C. § 107).

This decision builds on the Court’s earlier ruling regarding the occurrence of copyright-relevant reproductions based on the reproducibility of the content used for training. Even though the Court considers the recognisability of the training works in the outputs to be sufficient evidence of storage in the model, some key questions remain unanswered. In addition to the question of the required similarity and correspondence between the works, another key issue is how further influence by the users of the tool affects this assessment. The decision is not yet final, and it can be expected that further court rulings will be issued addressing these matters.

 

Background and functionality of Suno

The plaintiff is GEMA, a German collecting society that manages the rights of numerous music authors and music publishers. The defendant is the U.S.-based operator of the AI music generator Suno. GEMA claimed that several well-known musical works from its repertoire had been used without consent to train the Suno models and that these works subsequently appeared in the AI-generated output.

Suno allows users to create musical pieces based on prompts of varying complexity. According to the Court’s findings, the training datasets used for the models at issue were compiled from millions of audio recordings and metadata. Suno used the six musical works at issue in this process, including ‘Mambo No. 5’ by Lou Bega and ‘Forever Young’ by Alphaville. The Court found that the files had been downloaded from YouTube using so-called stream-ripping techniques, thereby circumventing YouTube’s protective measures against downloading the content.

To demonstrate the alleged infringements, largely identical prompts were used repeatedly, which essentially contained only the original lyrics, the title of the work, and a general description of the style. This resulted in the generation of audio files that showed significant similarity to the songs at issue. The number of iterations required to generate the output presented in court varied considerably depending on the work, ranging from four attempts for one song to 176 attempts for another. GEMA argued that the similarities proved that the musical works in question did not merely serve as the basis for statistical analyses, but had been ‘memorised’ within the model itself.

GEMA therefore filed a lawsuit seeking, among other things, an injunction and the disclosure of information to prepare a claim for damages, the amount of which could not yet be quantified.

 

Reproduction through the retrievable storage of training data

The Court largely ruled in favour of the claims brought by GEMA.

Key to the decision is the distinction made by the Court between the mere preparation of training data, the training of the model itself, and the subsequent generation of outputs. The Court first determined that the musical works at issue were indisputably part of the training dataset. However, the Court assumed that the songs were not merely used in the initial phase or that unprotected data derived from them was stored, but rather that the songs were actually embodied in the model’s memory. The Court defines ‘memorisation’ as a process in which a model does not merely learn information, patterns, or correlations from the training data, but rather encodes specific training content in the model parameters in such a way that this content can later be reproduced in whole or in part.

To prove such memorisation, the Court considers a comparison between the training data and the generated outputs to be sufficient. The decision was based on the fact that the prompts used by GEMA contained only the original song lyrics, the title of the work, and general stylistic guidelines, and did not provide specific instructions regarding melody, harmony, rhythm, or arrangement. The Court therefore describes these inputs as ‘simple, open-ended prompts’. It therefore regards the recognisability of the musical works found in the outputs as evidence that the works in question were embodied in the model and that it did not merely process abstract information about them.

On this basis, the Court concludes that the works at issue are embodied in the models and that this therefore constitutes a reproduction pursuant to Section 16 of the German Copyright Act (Urheberrechtsgesetz) (the ‘German Copyright Act’) and Article 2 of the Information Society Directive (Directive 2001/29/EC) (the ‘InfoSoc Directive’). The fact that the works were stored not in their original form but as numerical parameters was deemed irrelevant. The right of reproduction would also cover modern technical forms of storage. The sole decisive factor was that the works could be made perceptible again on the basis of the information contained in the model.

 

Exception for text and data mining not applicable

The Court then addressed in detail the question of whether these uses could be justified under the exception for text and data mining pursuant to Section 44b of the German Copyright Act, which is based on Article 4 of the Digital Single Market Directive (Directive (EU) 2019/790) (the ‘DSM Directive’). The Court first established that the provision may, in principle, also apply to the training of AI models and covers the preparatory acts of reproduction necessary for this purpose.

However, according to the Court, this exemption extends only as far as the extraction of information, patterns, or correlations from works, but ceases to apply where copyright-protected works themselves remain within the model. The Court once again distinguishes between reproductions necessary for the creation and preparation of the training dataset and those reproductions that persist in the model after training is complete. While the former could in principle be covered by Section 44b of the German Copyright Act, the elements of the works present in the model are no longer mere analysis results. The Court justifies this in particular with the wording of Article 4 of the DSM Directive and the respective German provision. The exception covers only reproductions made ‘for the purpose of text and data mining’. According to the Court, however, the reproductions in the model no longer served the purpose of further data analysis but rather enabled the subsequent reproduction of the works.

A broader interpretation would also conflict with the intent of the legislature. Permitting such use under the exception for the generation of such works would constitute a very significant encroachment on the exploitation interests of authors and rights holders. These interests must not be deprived of protection by an extension of the exception provision, which does not provide for any compensation for such exploitation.

The Court also rejected Suno’s argument that compliance with the obligations under Article 53(1)(c) and (d) of Regulation (EU) 2024/1689 (the ‘EU AI Act’), including the publication of a summary of the content used for training, would make further licensing unnecessary. Referring to Recital 108, the Court held that the EU AI Act does not alter the applicable EU copyright framework. Rather, the obligations under Article 53 complement the requirements under the DSM Directive and the InfoSoc Directive and do not provide an independent basis for the use of copyright-protected works.

 

Violation of right of communication to the public

In addition to the storage of the works in the model, the Court also found that the provision of the outputs generated by Suno constituted a copyright infringement. The chamber conducted a detailed musical analysis of the individual outputs and concluded that essential melodic, harmonic, and structural elements of the original works were recognisable. The Court therefore found that a communication to the public within the meaning of Section 15(2) of the German Copyright Act and Article 3 of the InfoSoc Directive occurred.

The Court also held the company liable for this and ruled that it could not claim to merely provide a neutral technical infrastructure. Rather, the generation of the outputs was based on the training dataset used by the company, the model architecture, and the training process carried out. As the Court considered the prompts used by GEMA to be merely simple, open-ended prompts, the generated outputs were attributable to the provider itself.

 

Violation of U.S. copyright law

Furthermore, the Court also examined the lawfulness of the training of the model, applying U.S. law to these actions. The subject of the review was the acts of reproduction carried out in the United States in connection with the procurement of training data and the training of the Suno models. The Court classified these acts as infringements of the right of reproduction under 17 U.S.C. § 106(1).

The Court also held that the use did not constitute a case of permitted fair use. In doing so, it addressed the case law of U.S. courts on this issue. While the Court acknowledged that the training of generative AI systems may, in principle, have a transformative character, in this specific case the circumstances weighing against fair use predominated. Unlike in the cases of Bartz v Anthropic and Kadrey v Meta, where training works were not recognisable in the outputs, the substantial similarity of the outputs to the originals negated any claim of transformative use. In addition, the Court took into account that, according to its findings, the training data had been obtained by circumventing YouTube’s technical protection measures.

In light of this, the Court ruled that the use was not justified under 17 U.S.C. § 107 and granted an injunction against the use of the works in question for training purposes in the United States.

 

Impact of the decision

The decision builds on the earlier ruling by the Court in the case of GEMA v OpenAI (Regional Court of Munich I, judgment of 11 November 2025, Case No. 42 O 14139/24). There too, the same chamber of the Court based its finding of reproduction on the fact that existing content could be reproduced upon request. That case concerned the memorisation of song lyrics that could be displayed upon entry of a prompt. Therefore, direct technical evidence demonstrating that a specific work can be concretely identified in the parameters of the model is not required, according to the Court. Instead, it can be sufficient to determine the existence of corresponding components of a work within the model based on the generated outputs. The ability to reconstruct the training works in a recognisable form thus becomes the key evidence of their storage within the model.

For operators of generative AI models in the EU, it may therefore be advisable to acquire the appropriate rights to use training data if the possibility cannot be ruled out that such data will reappear in a recognisable form in the output. Preventing the generation of outputs that resemble the works used for training in a manner relevant to copyright law may prevent evidence of use from being established. However, the requirements that apply in this regard and whether reproduction can be assumed even in cases where only individual parts of the training data appear in the output require further clarification. It also remains unclear how to handle cases in which the generated content is based on protected training data, yet the generated outputs deviate from the training works in a significant way.

The process of weighing and balancing the interests of the parties involved is still in its early stages, and numerous further decisions will be necessary to provide greater clarity in this area. Since this decision is not yet final, the Higher Regional Court of Munich or even the Federal Court of Justice may have to rule on this matter at a later stage. Although the Court did not consider it necessary to refer the matter to the European Court of Justice (ECJ) due to the allegedly clear guidelines, it is possible that such a referral may still occur in these proceedings, allowing the ECJ to comment on the interpretation of the underlying provisions of EU law.