GEMA v. Suno: What the Judgment Means for Generative AI and Copyright

The judgment of the Munich Regional Court I of July 31, 2026 in GEMA v. Suno is one of the most far-reaching German decisions to date on the application of copyright law to generative artificial intelligence. The Court addressed not only the generated musical outputs but the entire technical chain of use: the acquisition of training data, model training, the storage of protected works within the model, the generated outputs and the role of the platform operator.

The judgment is not yet final. An appeal is pending before the Munich Higher Regional Court under case number 6 U 2314/26 e.

Memorisation as Reproduction Within an AI Model

According to the Court, protected musical works may be reproduced within the AI model itself. The decisive issue is not whether a work is stored as a conventional audio file or in musical notation. A fixation may also exist within model parameters or other data structures if the work has been memorised in a manner that enables its recognisable reproduction.

The Court based this conclusion in particular on the fact that the musical works at issue could be reproduced in recognisable form using simple and open-ended prompts. The fact that a work is only indirectly perceptible within the model does not, in the Court’s view, preclude its classification as a reproduction.

This finding should not, however, be generalised to mean that every AI model trained on protected works automatically contains reproductions of those works. The judgment expressly depends on demonstrable memorisation and the reproducibility of the specific works concerned.

The Limits of the Text-and-Data-Mining Exception

The Court accepted in principle that Section 44b of the German Copyright Act and Article 4 of the DSM Directive may apply to reproductions made in preparation of an AI training corpus. It nevertheless distinguished between the creation and preparation of a dataset and the subsequent incorporation of protected expressive elements into a trained model.

In the Court’s view, memorisation of the musical works at issue went beyond the mere extraction of statistical information. It exceeded data analysis because the protected works were incorporated into the model parameters in a manner that allowed them to be reproduced. Such reproductions within the model were therefore not covered by Section 44b of the German Copyright Act.

The Court also found that the audio files had been acquired from YouTube using stream-ripping techniques that circumvented the platform’s so-called rolling cipher. It therefore concluded that there was no lawful access for the purposes of the text-and-data-mining exception.

Outputs and Communication to the Public

The outputs generated by the music generator were also relevant under copyright law. The Munich Regional Court I classified some of them as reproductions or adaptations of the protected musical works.

Particularly far-reaching is the Court’s view that the offering of a model may itself constitute a communication to the public where it provides access to memorised works through simple and open-ended prompts. According to the judgment, the protected work need not actually be retrieved by every user; making access possible may be sufficient.

This aspect is likely to receive particular attention on appeal. It partly shifts the copyright assessment away from the individual output and towards the model and its availability as such.

Responsibility of the AI Provider and Article 6 DSA

The Court held that Suno could not rely on the hosting safe harbour under Article 6 of the Digital Services Act. The outputs at issue were not merely third-party information supplied by users.

The prompts used in the case were simple and open-ended and did not determine the specific musical form of the outputs. The model and the application generated the music. The Court therefore classified the outputs as the provider’s own information.

This assessment may also be significant for other providers of generative systems. The less creative control users exercise and the more the content is determined by the model, the more difficult it may be to classify the resulting outputs exclusively as third-party user-generated content.

Training in the United States and Fair Use

According to the Court’s findings, the model had been trained in the United States. The Court therefore examined the reproductions made during training under US copyright law and rejected Suno’s reliance on fair use.

Its assessment took account of the commercial nature of the use, the highly creative character of the musical works, the use of complete recordings, the reproducibility of protected material in the outputs and the possible effects on licensing and sales markets. The acquisition of the training data through the circumvention of technical protection measures also weighed against the provider.

A German court’s interpretation does not bind courts in the United States. The judgment nevertheless illustrates that cross-border AI cases may require the parallel application and assessment of several national copyright systems.

Practical Significance of the Judgment

For providers of generative AI systems, the judgment points to several practical consequences.

The origin and lawful acquisition of training data should be documented. Technical measures against memorisation and near-identical reproduction should not be limited to filtering the final output. Rights reservations and licensing options must be incorporated into the data strategy. The respective roles of the user, the model and the platform in generating content should also be defined in a legally and technically transparent manner.

For right holders, the decision indicates which facts may be particularly relevant when enforcing claims: the reproducibility of a work through simple prompts, the composition of the training corpus, the origin of the training data and the technical functioning of the model.

Initial Assessment

The judgment adopts a strongly work- and right-holder-oriented approach. Its central doctrinal innovation lies in treating the model itself as containing a reproduction where sufficiently specific memorisation and reproducibility can be demonstrated.

The distinction between the extraction of statistical information and work-specific memorisation is persuasive and necessary. More open to debate are the conditions under which model parameters can qualify as a fixation of a protected work and the finding that merely offering the model may amount to a communication to the public.

The forthcoming appellate decision of the Munich Higher Regional Court is therefore likely to be important not only for AI music generators but for the copyright assessment of generative AI more broadly.

Decision

Munich Regional Court I, final judgment of July 31, 2026 – 42 O 763/25

Appeal: Munich Higher Regional Court – 6 U 2314/26 e