Germany

[DE] Munich Regional Court upholds GEMA’s claim against OpenAI for unauthorised reproduction of song lyrics

IRIS 2025-10:1/6

Christina Meese

Institute of European Media Law

In its judgement of 11 November 2025 (case no. 42 O 14139/24), the Landgericht München I (Munich Regional Court I – LG) ruled that the memorisation of linguistic works in AI language models, both when they were processed in the model and when they were delivered to the user in response to a corresponding prompt, constituted an act of reproduction within the meaning of copyright law. Although reproduction during the creation of training data material fell under the limitation of text and data mining, this did not apply to the process of training the model itself. As a consequence, the Munich Regional Court upheld a complaint filed by the GEMA collecting society concerning the processing of lyrics of its affiliated artists by ChatGPT. OpenAI, as the provider of ChatGPT, was accordingly ordered to refrain from these actions, to pay damages and to provide information about the scope of its use of the works and of the revenue generated from them.

GEMA had become aware, through its own sampling of ChatGPT (model 4 and user-defined agents based on model 4o), that the language model offered by OpenAI was able, when prompted, to reproduce song lyrics by GEMA artists, sometimes exactly and sometimes in a slightly modified form. These song lyrics were not freely available on the Internet, or at least not with the consent of the rightsholders. In addition, GEMA had generally claimed reservations of use as a limitation for text and data mining under Article 44b of the German Urheberrechtsgesetz (Copyright Act – UrhG), which transposed Article 4 of DSM Directive (EU) 2019/790 into German law. The collecting society had then brought an action before the Munich Regional Court I in relation to nine specific song lyrics retrieved from ChatGPT, including recent German hits such as “Atemlos” by Kristina Bach, older classics such as “Über den Wolken” by Reinhard Mey and songs written for special occasions such as “In der Weihnachtsbäckerei” and “Wie schön, dass du geboren bist” by Rolf Zuckowski. The court essentially upheld the claim.

According to the court, the memorisation (i.e. significant reproduction of training data) of linguistic works used by OpenAI constituted reproduction within the meaning of Article 16 UrhG (which transposed Article 2 of InfoSoc Directive 2001/29/EC). The lyrics were (1.) physically fixed in the models because the lyrics that had served as training data were reproducibly contained in the model and thus embodied. In particular, the comparison between the original work and the output of a simple prompt (e.g. “reproduce the chorus of the song ‘Atemlos’”) submitted by GEMA was sufficient to convince the court that the work at issue had been memorised – even without knowledge of the specific training data used to develop ChatGPT. In addition, (2.) the linguistic works could be made indirectly perceptible via corresponding user interfaces. The court did not accept OpenAI’s objection that ChatGPT essentially only strung together the most likely words and that responses to prompts were therefore not always identical.

 

Furthermore, the reproduction that took place in the models was not covered by the text and data mining limitation. Although such language models would, in principle, fall within the scope of the limitation, this only applied to the “pre-training phase” during which the data corpus was compiled for training, i.e. crawled data was converted into machine-readable text. However, it did not cover the subsequent training phase, in which information was extracted from the data corpus and works were reproduced, since this was not done in preparation for text and data mining. With regard to the interpretation of Article 4 of the DSM Directive, the court stated in particular that: “A presumably technology- and innovation-friendly interpretation that also considers reproductions in the model to be covered by the limitation is prohibited in view of the clear wording” (para. 208).

The court also considered further limitations irrelevant. In particular, implied consent on the part of rightsholders could not be taken into account because the training of language models could not be considered a common and expected type of use that rightsholders should anticipate.

Finally, the court held that the operators of the language model were also liable for copyright infringements committed by outputs because they had control over the process. It was true that such control could be lost to the user if outputs were “provoked” by the user. However, this was not the case with simple prompts, as was the case here.

OpenAI was therefore ordered to refrain from reproducing the nine song lyrics at issue, both in the model and in outputs. In the event of non-compliance, a fine of up to €250 000 or, alternatively, imprisonment could be imposed in each case. In addition, OpenAI was ordered to provide information on the extent to which the reproductions in dispute had been made and how much revenue had been generated from them. OpenAI was also obliged to compensate GEMA for the damage it had already suffered and would suffer in future as a result of the copyright infringements. The only element of the complaint that was dismissed was the infringement of general personality rights also alleged by GEMA in connection with the incorrect attribution of modified song lyrics to authors.

The Munich Regional Court I did not consider it necessary to refer the case to the European Court of Justice (ECJ). The case law on the ECJ’s broad interpretation of acts of reproduction left no doubt as to its application to language models. The text and data mining limitation was so clearly inapplicable that a referral on this basis was also unnecessary. In addition, the court referred to the pending case C-250/25 from Hungary, in which the ECJ would clarify these questions of interpretation.

Although the decision only relates to the nine song lyrics at issue, it sends out a strong signal. However, this first-instance judgement is not yet final and it is very likely that OpenAI will appeal. At the same time, GEMA has a case pending against Suno before the same court chamber concerning compositions. It has already made it clear that it wants to claim licence fees for both training and reproduction of outputs, as well as for the use of outputs by users, for example by making them available to the public.