AI and Copyright: Munich Regional Court convicts AI music generator Suno – including under US law

‍1. Background

Whether, and under what conditions, training AI models on copyright-protected works is permissible is one of the most important unresolved questions of copyright law. German courts have been confronted with these issues in three landmark proceedings, each presenting a different factual configuration: the LAION proceedings before the Hamburg Regional and Higher Regional Courts (case no. 5 U 104/24 at the OLG, currently I ZR 281/25 at the BGH), the GEMA v. OpenAI proceedings before the Munich Regional Court (case no. 42 O 14139/24, currently 6 U 3662/25 at the OLG Munich), and the GEMA v. Suno proceedings discussed here, also before the Munich Regional Court.

The proceedings concern different factual constellations, but all share a common conceptual framework: the three-phase model, which traces back to Spindler (GRUR 2016, 1112) and was already adopted by the Hamburg Regional Court in the LAION proceedings. Phase 1 covers the extraction and preparation of training data; Phase 2 covers the actual training of the model; Phase 3 covers the use of the trained model through prompts and outputs.

The Hamburg LAION proceedings concerned the creation of a non-commercial image-text dataset, i.e. the first phase of data collection. The OpenAI proceedings concerned a claim by GEMA against a large language model that reproduced protected song lyrics as output. The focus there was on Phase 2 – the training of the model through analysis of data material and its enrichment with metadata. The Munich Regional Court held in those proceedings that the “memorisation” of the lyrics in the model, which was then able to reproduce them identically in response to appropriate prompts, constituted an impermissible reproduction of the protected work.

The phenomenon of so-called memorisation can be explained as follows: when training an AI model, training data is analysed and statistical patterns are transferred into the model parameters. In the – from the provider’s perspective – ideal case, the model learns abstract relationships such as grammar, style or musical conventions. In certain cases, however, the absorption can go beyond mere patterns: the model then stores sufficient information in its parameters to reproduce a specific training work, in whole or in part, identically. This state of “learning by heart” is referred to in information technology research as memorisation. The word-for-word or near-identical reproduction of memorised content in the output is then described as “regurgitation”.

The present Suno proceedings concern, like the OpenAI proceedings, Phase 2 – here involving a claim against an AI music generator that produces complete musical pieces from text prompts and in doing so reproduces the melodies, harmonies and arrangements of the training works in a recognisable manner.

2. Facts

GEMA brought proceedings against the US-based operator of the AI music generator Suno for infringement of copyright in six musical works, including “Atemlos durch die Nacht”, “Big in Japan” and “Forever Young”.

The defendant trained its system, which is based on transformer and diffusion models, using millions of complete audio recordings that it extracted from YouTube by means of stream-ripping in circumvention of the so-called “Rolling Cipher” – a technical protection measure designed to prevent the downloading of content. The model itself is made available to the German market on edge servers located in Germany.

GEMA’s staff were able to generate outputs closely resembling the original works by entering the original lyrics together with general genre indications (such as “Schlager” or “80s, Synth pop”) and the titles of the works. They provided no instructions relating to melody, harmony, rhythm or arrangement. Between 4 and 176 identical prompts per work were required before a recognisable output was produced.

3. The decision

The 42nd Civil Chamber granted the claim in large part and ordered Suno to cease and desist, to provide information and to pay damages – both in respect of the reproductions in the model and in the outputs under German law, and in respect of the training reproductions in the United States under US law. The latter is remarkable and, so far as can be seen, unique in the rejection of the fair use defence by a German court.

a) Local and international jurisdiction under § 131 of the Collecting Societies Act (VGG)

A novel aspect of the decision concerns local and international jurisdiction. Suno is domiciled in the United States, and the training acts undisputedly took place there. Nevertheless, the court affirmed the jurisdiction of the Munich Regional Court over the infringements in the United States as well.

For the domestic acts, jurisdiction derives from § 131(1) sentence 1 VGG in conjunction with § 32 ZPO: GEMA as a collecting society may sue at the place of the infringing act, which also falls within the jurisdiction of the Munich Regional Court, since the model is hosted on servers in Germany and generates infringing outputs throughout Germany.

For the training acts in the United States, the court then relies on the consolidation rule in § 131(2) VGG: where several sets of proceedings brought by a collecting society against the same infringer fall within the jurisdiction of different courts, the collecting society may consolidate all claims before one of those courts. The court applies this provision with double function, relying on it to establish not only local but also international jurisdiction. While this is correct in principle, it leads to the questionable outcome that the jurisdiction of the German courts is established on an originary basis through the consolidation rule, even though the German courts would not otherwise have jurisdiction over the infringements in the United States.

This construction is not entirely persuasive, and whether it will hold will ultimately be decided by the BGH. For the time being, this jurisdictional basis carries significant signal effect: German collecting societies could, on this analysis, consolidate all claims against a globally operating AI provider before a German court, even if the infringing training acts took place exclusively abroad.

b) Memorisation and reproduction in the model

On the central question of whether the musical works in dispute were reproduced in the model, the court follows the same methodology as in the OpenAI proceedings. It found memorisation proven by comparing the works that were undisputedly used as training data with the outputs, pursuant to § 286 ZPO.

In this context, a closer examination of the prompts – and in particular the question of whether entering the lyrics determines the output musically – is decisive. The court answers this question in the negative, with the following detailed reasoning: a text sets at most a thematic framework and requires prosodic compatibility. It does not, however, determine the specific melodic sequence, the rhythm, the harmonic structure or the tempo. The same text can be set to music with entirely different melodies. Genre indications such as “Schlager” or “Disco” are too general to fix the music. The prompts are therefore to be characterised as simple and open-ended – not as prompts that provoke or steer the output.

The fact that GEMA submitted between 4 and 176 identical prompts per work does not alter this conclusion: since the prompt remains unchanged, the probability distribution of the model also remains fundamentally the same. The repeated input does not steer the output in any particular direction.

The court otherwise confirms the line taken in the OpenAI proceedings: Art. 2 of the InfoSoc Directive captures, in a technologically neutral manner, every reproduction “by any means and in any form”. Decomposition into parameters does not preclude the existence of a physical fixation; a concretely identifiable data storage in the model is not required.

c) The TDM exception and the AI Act

On the TDM exception, too, the court follows the line of the OpenAI judgment: § 44b Copyright Act covers reproductions made in assembling the training corpus (Phase 1), but not the further memorisations in the model (Phase 2). Memorisations serve no further data analysis and therefore exceed the purpose of text and data mining.

The decision has independent significance in respect of two further points:

First, the court denies that lawful access existed under § 44b(2) Copyright Act. Suno had obtained its training data by stream-ripping in circumvention of the Rolling Cipher on YouTube. In the court’s view – in line with a decision of the Hamburg Higher Regional Court (5 U 54/23) – this protection mechanism constitutes an effective technological measure under § 95a(2) Copyright Act. Its circumvention therefore precludes lawful access. This aspect did not play a comparable role in the OpenAI proceedings, since the provenance of the training data was not in dispute in the same way there.

Second, the court addresses Art. 53(1)(c) and (d) of the AI Act in detail. Suno had argued that, by producing a training data summary and a copyright strategy in accordance with the AI Act, it had done everything required to comply with copyright law. The court rejects this: Art. 53 AI Act does not define the scope of the DSM Directive, but refers to it. The summary serves to facilitate enforcement of rights, not to legitimise copyright infringement. The Code of Practice on the AI Act also makes clear that adherence to it does not constitute compliance with copyright law.

For the overall assessment of this question, it should be noted that the OLG Hamburg affirmed the general applicability of § 44b Copyright Act to AI training in the LAION proceedings. There is, however, no contradiction: the decisive difference is that in the Munich proceedings memorisations were established – that is, reproductions that go beyond mere information extraction in Phase 1 and encroach upon the exploitation right.

d) Communication to the public and attribution to Suno

On the right of communication to the public, the court takes a different approach from the OpenAI proceedings. In those proceedings the court had affirmed § 19a Copyright Act; here it denies the making available to the public: GEMA had not sufficiently demonstrated that the works in dispute could be accessed with any prompt. Where up to 176 identical prompts are required, access does not occur at a time chosen by the user.

Instead, the court affirms an infringement of the unnamed right of public communication asserted in the alternative under § 15(2) sentence 1 Copyright Act. Making the model available with its memorised works through the user interface opens up the possibility of experiencing the works; the defendant acts directly and not merely as an intermediary. Even the mere possibility of access constitutes an act of communication, since this right attaches to the act of making available, not to the specific act of retrieval.

As regards the attribution of the output reproductions, the court characterises Suno as a direct perpetrator. The defendant exercises control over the act: it selected the training data, determined the model architecture and conducted the training. The simple, open-ended prompts submitted by GEMA’s staff did not manipulate the output, and the mere triggering of a reproduction by entering a prompt does not make the user the party responsible for the reproduction. Suno is therefore comparable neither to an internet radio recorder (cf. BGH, judgment of 27 June 2024, I ZR 14/21 – Internet-Radiorecorder II) nor to a sales platform (cf. BGH, judgment of 23 October 2024, I ZR 112/23 – Manhattan Bridge).

The court finally also denies the liability privilege under Art. 6 DSA: the outputs are the defendant’s own content, not information provided by users. Suno does not merely transmit user-provided content, but plays an active role through training and model architecture.

e) Fair use under US law

The most spectacular part of the decision concerns the application of US law. For the reproduction acts carried out in the United States for training purposes, the court was required to apply US law pursuant to the lex loci protectionis principle (Art. 8(1) of the Rome II Regulation). The court ascertains the applicable US law of its own motion under § 293 ZPO, drawing on, among other sources, the judicial decisions submitted by the parties and the library of the Max Planck Institute for Innovation and Competition.

The focus is on the fair use doctrine of 17 U.S.C. § 107, which examines, by reference to four factors, whether an unlicensed use of a work is exceptionally permissible. The court explicitly distances itself from the only two US decisions to date on AI training in which the courts affirmed fair use: Bartz v. Anthropic and Kadrey v. Meta.

The decisive difference, in the view of the Munich Regional Court: in both US proceedings, substantially similar outputs were absent. The courts in those cases expressly emphasised that their assessment would differ if the outputs reproduced the training works. That is precisely the situation in the Suno proceedings.

On the first factor (purpose and character of the use), the court applies the standard set by the Supreme Court in Andy Warhol Foundation v. Goldsmith (2023): what matters is not merely whether the use is abstractly transformative, but whether the specific use is justified. The defendant’s outputs served the same purpose as the originals – they constituted substantially similar musical works that could be listened to. Moreover, Suno acted commercially and had obtained the training data by circumventing the Rolling Cipher, which the court treats as bad-faith acquisition within the meaning of the Supreme Court’s jurisprudence in Harper & Row v. Nation Enterprises.

On the second factor (nature of the copyrighted work), the creative works at the heart of copyright protection weigh against fair use. On the third factor (amount and substantiality of the portion used), the full reproductions are not to be characterised as mere intermediate copies, since they manifest in substantially similar outputs. On the fourth factor (effect on the market), the outputs constitute a substitute product: the creation of cover versions using the music generator is a well-known and documented pattern of user behaviour, as evidenced, for example, by YouTube tutorials that explain step by step how users can create cover versions of their favourite songs with Suno.

The judgment grants GEMA a permanent injunction under 17 U.S.C. § 502(a), applying the eBay standard. The court expressly considers itself capable of applying the US equitable standard (equity), since German courts also regularly apply considerations of equity, for example through § 242 BGB or § 139(1) sentence 3 of the Patents Act.

For the disclosure claim under US law, the court resolves a conflict-of-laws gap: US law recognises rights to information only as a procedural discovery mechanism, while German procedural law has no equivalent pre-trial discovery. The court remedies this by applying § 242 BGB by analogy.

4. Assessment and outlook

The decision, which runs to well over one hundred pages, reinforces the line of the 42nd Civil Chamber of the Munich Regional Court on the copyright assessment of AI models and develops it in several respects.

On memorisation, the TDM exception and the attribution of output reproductions to the model provider, the key reasoning of the OpenAI and Suno judgments is consistent. The two Munich judgments can in turn stand without contradiction alongside the Hamburg LAION judgment: the OLG Hamburg affirmed the general applicability of the TDM exception to AI training – but in a case in which only the preparatory data collection (Phase 1) was at issue and memorisation was neither established nor in dispute.

An interesting and persuasively reasoned difference between the two Munich judgments concerns the right of communication to the public. In the OpenAI proceedings the court still affirmed § 19a Copyright Act, whereas in the Suno proceedings it denied it, on the ground that access at a time of the user’s choice is not assured where numerous prompts are required to trigger regurgitation. Instead, the unnamed right of public communication under § 15(2) Copyright Act applies.

The jurisdictional reasoning under § 131(2) VGG is likely to attract the most controversy. The court derived its local and (thereby) international jurisdiction over the training acts in the United States from the fact that the provision permits, in favour of collecting societies, a consolidation of jurisdiction over multiple claims against the same infringer – including where the court would not otherwise have jurisdiction over the extraterritorial infringing act at all. In other words, a German court is to have jurisdiction over acts carried out on US territory (here: the training of the model) where a claim is simultaneously brought that involves an infringing act in Germany (here: the memorisation on servers located in Germany). Whether this far-reaching interpretation will hold is open to question and will ultimately be resolved by the BGH. If it does hold, however, it would open up the possibility for German collecting societies to pursue purely foreign matters before German courts and would significantly strengthen their means of enforcement against US-based AI providers.

On the merits, however, the training in such cases falls to be assessed under the law of the place of training, here US law. The fair use analysis by a German court is a novelty. It is likely to attract interest in the United States as well, and may be drawn on as a reference point in the US discourse. A subsidiary point that will nonetheless play a role as the proceedings continue: the application of foreign law is not, as such, subject to review on appeal on a point of law (§§ 545(1), 560 ZPO), with the result that the Munich Higher Regional Court is the final instance on this question.

On the merits, the court confirms its position: whoever memorises copyright-protected works in a model and reproduces them in the form of substantially similar outputs reproduces those works – regardless of whether the reproduction is effected in model parameters or on a conventional storage medium. A business model premised on the appropriation of others’ intellectual property free of charge finds no support in either German or US law.

The decision is not yet final. The BGH will at the latest have to rule on the central questions of German law and will in doing so be hard pressed to avoid referring questions on Art. 2 of the InfoSoc Directive and Art. 4 of the DSM Directive to the CJEU. A preliminary reference on related questions is already pending at the CJEU (Case C-250/25 – Like Company).

tl;dr: The Munich Regional Court has held the operator of the AI music generator Suno liable for infringement of copyright in six musical works. Memorised musical works in AI models constitute reproductions that are not covered by the TDM exception in § 44b Copyright Act. Circumvention of the Rolling Cipher during data acquisition also precludes lawful access. Under § 131(2) VGG, a German court with jurisdiction over the copyright assessment of reproductions in the model may also consolidate claims in respect of training acts carried out on US territory. The US fair use doctrine is unavailable where the outputs substantially reproduce the training works.