Ruling of the Munich Regional Court (GEMA v OpenAI) on copyright and AI training
In the ruling on November 11, 2025, the 42nd Civil Chamber of the Munich I Regional Court[1], which is specialised in copyright law, upheld the claims for injunctive relief, information and damages asserted by GEMA against two companies of the Open AI group.
In this court proceeding, Germany’s music collecting society GEMA brought an action for infringement of copyright before the Munich I Regional Court (case no. 42 O 14139/24) against two companies of the Open AI group. The infringement concerns the lyrics of nine well-known German authors (including ‘Atemlos’ by Kristina Bach and ‘Wie schön, dass du geboren bist’ by Rolf Zuckowski). The plaintiff argued that the song lyrics were memorised in the defendant’s language models and, when the chatbot was used, were outputs largely unchanged in response to simple user queries.
The defendants are operators of language models and chatbots based on them. They had objected to the claims, arguing that their language models did not store or copy specific training data, but rather reflected in their parameters what they had learned based on the entire training data set. Since the outputs were only generated as a result of user input (prompts), it was not the defendants but the respective user who was responsible for the output as its creator. In any case, any legal infringements were covered by the limitations of copyright, in particular the limitation for so-called text and data mining.
According to the decision of the Court, the plaintiff is entitled to the asserted claims both based on the reproduction of the texts in the language models and their reproduction in the outputs. Both the memorisation in the language models and the reproduction of the song lyrics in the chatbot’s outputs constitute infringements of copyright exploitation rights. These are not covered by any restrictions, in particular the restriction on text and data mining.
In the opinion of the Chamber, the song lyrics in dispute are reproducible in the defendant’s language models 4 and 4o. It is known from information technology research that training data can be contained in language models and can be extracted as outputs. This is referred to as memorisation. This occurs when the language models not only extract information from the training data set during training, but also completely adopt the training data in the parameters specified after training. Such memorisation was determined by comparing the song lyrics contained in the training data with the reproductions in the outputs.
Reproduction right
The memorization of the disputed song lyrics in the models constitutes reproduction in accordance with § 16 of the German Copyright Act (UrhG)[2]. The reproduction right regulated in § 16 UrhG implements Article 2 of the InfoSoc Directive[3]. Recital 21 of the InfoSoc Directive already requires a broad definition of reproduction in order to guarantee legal certainty in the internal market. This is reflected in the wording of Article 2 of the InfoSoc Directive, which uses all-encompassing terms such as “direct or indirect,” “temporary or permanent,” and “in any manner and in any form” to describe reproduction.
Applying these standards, the models constitute reproductions. The texts in dispute are physically embedded in the models and can be made indirectly perceptible. Due to the memorization process, embodiment is a prerequisite for the reproduction of the song lyrics in question through data within the specified parameters of the model. The song lyrics in question are reproducibly fixed in the models.
New technologies, such as models, are covered by the reproduction right under Article 2 of the InfoSoc Directive and §16 of the German UrhG. According to Recital 5 of the InfoSoc Directive, the reproduction right regulated in the InfoSoc Directive is intended to adequately take account of the new forms of exploitation resulting from technological developments. With regard to the song lyrics in dispute, which were used as training data, it is sufficient if the model is able, based on the statistical information derived in the training, to generate statistically probable token sequences that recognizably represent the song lyrics.
The song lyrics in dispute, which are physically defined in the models, can be made indirectly perceptible to the human senses. According to the case law of the Court of Justice of the European Union, indirect perceptibility is sufficient for reproduction, which is given if the work can be perceived using technical aids
This reproduction in the models is not covered by the limitation provisions for text and data mining in § 44b of the German UrhG or by §57 UrhG as an insignificant accessory. Language models are, in principle, covered by the scope of the text and data mining restrictions. According to § 44b of UrhG, text and data mining is the automated analysis of single or multiple digital or digitized works to extract information, particularly about patterns, trends, and correlations. The limitation in the UrhG and its European legal basis in Article 4 of the DSM Directive[4] applies to text and data mining in the training of artificial intelligence. The provisions cover necessary reproductions when compiling the data corpus for training, such as the reproduction of a work by converting it into another (digital) format or storing it in the working memory. The reasoning behind this is that these reproductions are only made for subsequent analysis purposes and therefore do not affect the author’s exploitation interests in the work. Since these purely preparatory acts for text and data mining do not affect any exploitation interests, the law does not provide for any obligation to pay remuneration to the author.
If, as in this case, not only is information extracted from training data during training, but works are also reproduced, this does not constitute text and data mining in the opinion of the Chamber. The premise of text and data mining and the related restrictions, namely that the automated evaluation of mere information does not affect exploitation interests, does not apply in this constellation. On the contrary, the reproductions in the model interfere with the exploitation rights of the rights holders.
Another interpretation, presumably favourable to technology and innovation, which also sought to consider reproductions in the model as covered by the restriction, is precluded by the clear wording of the provision. An analogous application is also out of the question.
Even if one were to assume an unintended regulatory gap because the legislator was not aware of memorisation and the associated permanent reproduction in the models that is relevant under copyright law, there is no comparable interest. The limitation provision regulates the permissibility of preparatory acts of reproduction in text and data mining, a situation in which the exploitation interests of the authors are not jeopardised because mere information is extracted and the work as such is not reproduced. In the case of reproductions in the model, however, the exploitation of the work is permanently impaired and the legitimate interests of the rights holders are thereby infringed. The authors and rights holders would thus be left unprotected by an analogous application of the limitation provision, which does not provide for remuneration for exploitation. The risk of memorisation stems solely from the sphere of the defendant. In the case of an analogy of the limitation, only the infringed rights holder would bear this risk.
In the absence of a main work, the reproductions of the song lyrics at issue did not constitute an inadmissible accessory under Section 57 of the German Copyright Act (UrhG). Contrary to the defendant’s view, the song lyrics could not be regarded as incidental and dispensable alongside the entire training data set. This would require that the entire training data set also be a copyright-protected work.
The defendants’ interference with the plaintiff’s exploitation rights was also not justified by the consent of the rights holders, as the training of models could not be considered a normal and expected type of use that the rights holder would have to anticipate.
Even by reproducing the song lyrics in the chatbot’s outputs, the defendants had, according to the Chamber’s decision, unlawfully reproduced and made publicly available the song lyrics at issue. The original elements of the song lyrics were always recognisable in the outputs.
The defendants, and not the users, were responsible for this. The outputs were generated by simple prompts. The defendants operated the language models for which the song lyrics were selected as training data and with which they were trained. They were responsible for the architecture of the models and the memorisation of the training data. Thus, the language models operated by the defendants had a significant influence on the outputs, and the specific content of the outputs was generated by the language models.
The interference with the exploitation rights by the outputs is also not covered by a limitation provision.
The judgment is not yet final.[5]