GEMA v Suno AI: key EU decision on AI training, outputs and copyright

On 31 July 2026, the Munich District Court handed down their judgment in GEMA v Suno AI (42 O 763/25), largely upholding the claims of GEMA, the German music collecting society and CMO. The case marks the second success in the past year for GEMA, following the Regional Court's November 2025 decision in GEMA v OpenAI (42 O 14139/24) (although this is under appeal).

The judgment follows a claim filed on 21 January 2025, in which GEMA sought an injunction, disclosure, and damages against Suno AI on the grounds of infringement of copyright regarding the composition of six musical works ¹ (but not their sound recording) in relation to their use in training an AI model and for their alleged reproduction in outputs.

¹ Atemlos durch die Nacht, Big in Japan, Daddy Cool, Forever Young, Rasputin, and the refrain of Mambo No. 5.

Background

Suno AI develops, operates and licences an AI-based music generator, with their model being stored on servers in Germany. Outputs generated in response to prompts by GEMA in versions 3.5 and 4 of the model were alleged to infringe the musical works in question.

Much of the factual background was undisputed. Suno used the works at issue as training data for its model, notwithstanding the general reservation of rights notice, and notices placed on YouTube. These works were obtained via YouTube using "stream-ripping" techniques, thereby circumventing the platform's technical protection measures (the "rolling cipher"). As part of the training process, which occurred in the US, Suno linked the metadata (including title, genre, and lyrics) to the audio files, creating "bundles" which were tokenised and used to train the model. This training data was stored for security and backup purposes.

Within their evidence, GEMA submitted outputs from the model based on 'simple prompts' which named only the title, lyrics, and musical style of the respective works. Identical prompts were repeated between 4 and 176 times, with the outputs that were most strikingly similar to the relevant works being documented to demonstrate a pattern of consistent infringement.

Issues

Memorisation and reproduction

Comparing the outputs generated on Suno's platform against the relevant works, the court found that the training data had been 'memorised'. The works were capable of being reproduced by 'simple, open-ended prompts' (rather than controlled, provocative prompts), to such an extent that there was 'no doubt as to the recognisability of the original musical works in the outputs'. Further, storage of the training data on German servers (even if as temporary or statistically probable token sequences that recognisably reproduce musical works embodied in the model) was found to infringe the reproduction right under s.16(1) of the German Copyright Act ("UrhG") (implementing Article 2 of the InfoSoc Directive). The Court recognised a test of "indirect perceptibility" (through a music generator) as sufficient for reproduction.

The court also rejected Suno's attempt to rely on the text and data mining exception (the "TDM"), as 'reproductions within the model do not serve the purpose of further data analysis'; the exception only applies to extraction of data, not to the subsequent reproduction of works. In any case, the relevant works were not obtained lawfully by Suno (given Suno's circumvention of the rolling cypher mechanism), undermining any reliance on the TDM under s.44(b) UrhG.

Attribution and communication

Suno attempted to argue that responsibility for the infringing outputs 'lies solely with the claimant as the user', and that GEMA's prompts were 'complex' and 'deliberately aimed at infringing copyright'; however, this argument was rejected by the court. The infringing outputs were 'attributable to systemic causes' including the model's training data and architecture, not GEMA's prompts, which were 'simple' and 'open-ended'. The defendant, being responsible for the architecture of the models and offering the models for use by users, was the perpetrator, with control over the act of reproduction through the outputs.

It was further held that the outputs constituted infringement of the 'unnamed right to communication to the public' (s.15(2) UrhG). Considering the existence of infringing outputs and the reach of Suno's platform, public access was fulfilled from both a quantitative and qualitative perspective. The works were made available for retrieval in a recognisable form, to a potentially large number of people and to a new audience. The mere possibility of access constitutes a relevant communication to the public.

Intermediary

The Court rejected Suno's reliance on Article 6 of the EU Digital Services Act ("DSA"), as the defendant did not merely provide a technical, content-neutral infrastructure platform, but as an operator played an active role in offering the model for use and in generating outputs. Accordingly, Suno was not acting simply as an intermediary which stored information.

Jurisdiction

The Munich court derived international and territorial jurisdiction over the alleged infringing acts through s.131(1) of the German Collecting Societies Act (the "VGG") and s.32 of the German Code of Civil Procedure (the "ZPO"). In addition, s.131(2) of the VGG allows a collecting society to consolidate claims against a single infringer before a particular court, where those claims lie across different jurisdictions. The court were therefore able to exercise jurisdiction over the US-based training through the principle of dual functionality.

Fair use

Regarding the training which occurred in the US, the court applied a detailed analysis of US law to find infringement by Suno. On all four factors assessed under fair use (17 U.S.C. § 107), the court 'weighed in favour of the composers', dismissing Suno's defence and distinguishing the present case from recent US trial decisions (Bartz v Anthropic and Kadrey v Meta). Significantly, it was demonstrated that the infringing outputs were 'substantially similar to the original works', so were therefore capable of causing market harm to GEMA (and its members).

EU AI Act

Suno attempted to assert that compliance with the EU AI Act, specifically Article 53(1), extended to compliance with copyright law and replaced any requirement for licensing. This argument was firmly rejected by the court, who recognised that such an interpretation would leave rights holders 'without recourse'. The court upheld the legislative purpose of the Act and its accompanying guidance, which confirms that 'adherence to the Code does not constitute compliance with the Union law on copyright and related rights'.

Transient copying exception

The Court rejected any reliance by Suno on a defence of transient or incidental copies on the basis that the reproductions or storage of the works were of independent economic significance. In addition, they fell outside the list of permitted purposes, such as transmission within a network between third parties, via intermediaries. 

Outcome

GEMA was granted injunctive relief prohibiting: (i) reproduction for training purposes; (ii) reproduction within the model; (iii) public performance through the offering by Suno; and (iv) infringements relating to distribution. Notably, the injunction extended extra-territorially, as Suno were ordered to refrain from using the relevant works for training purposes 'within the territory of the United States of America'.

Suno has also been ordered to disclose the extent of its infringing acts and to pay damages, although these have not yet been quantified. In cases such as the present, it was discussed that damages may be difficult to quantify, as the extent of infringement and harm caused is indeterminate.

The judgment is not final, and an appeal may be brought. In parallel, the pending case in the CJEU (Like Company v Google, C-250/25) may provide more clarity in relation to the TDM exception and reproduction, with an opinion expected from the Advocate-General on 3 September 2026.

Implications

This judgment has been issued against a backdrop of ongoing litigation and also take-up of licensing between rightsholders and AI model providers. No less than two weeks following the judgment, Suno AI have announced a licensing deal with German music rights manager BMG. This follows GEMA's launch of 'PLAI' a week prior to the Judgment, offering licensed, curated, and scalable datasets specifically for the purpose of training AI tools.

On the face of this very detailed and reasoned decision (and the decision in GEMA v OpenAI), clear infringement has been found in terms of training the model and in the outputs through memorisation in the parameters of the model. This differs from the finding in Getty Images v Stability AI in the English High Court, which found no memorisation in that particular model, and also no secondary copyright infringement through the import of the model as an "infringing copy" (as defined in ss. 22 and 23 of the UK Copyright, Designs and Patents Act 1988) (though this case is on appeal to be heard by the English Court of Appeal on 10 November 2026).

At present, we are left with potential differences in approach between the UK and EU courts (subject to any appeals) over memorisation, notwithstanding a recognition that outcomes may be highly fact-specific in regard to the particular model, its training process, and how it generates outputs. This means that expert evidence will be critical to the way that AI developers present their case, as well as how rightsholders generate evidence of repeatable outputs, without the results being over-engineered and suggestive. As with US litigation, disclosure (discovery) of critical internal documents may play a critical role.