In recent years, the development of artificial intelligence (AI) has accelerated due to increased data availability and the use of multilayer artificial neural networks. There is a tension with copyright law because large language models, such as GPT-4, are trained with copyrighted texts and can reproduce them.
On November 13, 2024, the German collecting society GEMA filed a lawsuit against OpenAI, LLC, and OpenAI Ireland, Ltd. (hereafter referred to together as “OpenAI”) in the Munich Regional Court (case number 42 O 14139/24). The case concerns ChatGPT’s reproduction of song lyrics. GEMA argues that, in response to simple prompts, ChatGPT reproduces the song lyrics on which it appears to have been trained. One year later, on November 11, 2025 the Munich Regional Court announced its decision. The text of the decision can be found online (https://www.gesetze-bayern.de/Content/Document/Y-300-Z-GRURRS-B-2025-N-30204?hl=true ).
When simple prompts are entered, the OpenAI chatbot ChatGPT reproduces the lyrics of the songs. OpenAI had argued that the outputs were based on a sequential-analytical, iterative-probabilistic synthesis and represented an independent and original (machine) generation of its own kind.
The Munich Regional Court has prohibited OpenAI from, among other things,
- reproducing the song lyrics, in whole or in part, in large language models without the plaintiff’s consent and/or having this action performed by third parties, as occurred in models 4 and 4o.
- making the song lyrics, in whole or in part and/or in the form of adaptations, publicly available in the outputs of a generative language assistant (chatbot) without the plaintiff’s consent, and/or reproducing them and/or having these actions carried out by third parties, as occurred.
OpenAI was also ordered to provide information and pay damages.
The Munich Regional Court thoroughly examined the technical and legal basis. The judgment begins with a description of how artificial neural networks and the transformer architecture underlying the GPT series models operate. In the legal analysis, the court determines the applicability of German law. For the copyright infringements, against which GEMA seeks protection for the territory of the Federal Republic of Germany, the applicability of German law is based on Art. 8 (1) of the Rome II Regulation (https://eur-lex.europa.eu/eli/reg/2007/864/oj ).
According to Art. 8 (1) of the Rome II Regulation, non-contractual obligations arising from an infringement of intellectual property rights are governed by the law of the country for which protection is sought.
The Munich Regional Court ruled that the song lyrics were reproduced in the language models. The lyrics in question have been memorized by the models. According to the court, memorization can be determined by comparing the lyrics with the outputs. The use of the disputed song lyrics as training data is undisputed. The disputed lyrics are easily identifiable with simple prompts such as “What are the lyrics to [song title],” “Who wrote the lyrics,” “What is the chorus of [song title]”, “Please also tell me the first verse”, and “Please also tell me the second verse”. According to the Regional Court Munich’s assessment, the lyrics in question are physically fixed in the models and can be indirectly perceived. Due to memorization, embodiment is a prerequisite for the copyrighted reproduction of the disputed song lyrics by data within the model’s specified parameters. The song lyrics in dispute are reproducibly fixed in the models. The Munich Regional Court refers to Art. 2 of the InfoSoc Directive. According to Art. 2 of the InfoSoc Directive, the reproduction right covers reproductions in any manner and in any form (https://eur-lex.europa.eu/eli/dir/2001/29/oj ).
According to the Munich Regional Court, the exceptions for text and data mining in Sections 44b and 60d of the German Copyright Act do not permit these reproductions ( https://www.gesetze-im-internet.de/englisch_urhg/englisch_urhg.html#p0328 ). These German provisions implement the corresponding provisions of the DSM Directive (https://eur-lex.europa.eu/eli/dir/2019/790/oj ).
According to the Munich Regional Court, reproducing the song lyrics in question in the models does not constitute text and data mining. Text and data mining aims to generate information. However, in the opinion of the Munich Regional Court, the memorization of the lyrics exceeds this generation of information and is therefore not merely text and data mining. The song lyrics, used as training data, were fully integrated into the model’s parameters, which interferes with the authors’ exploitation interests.
Art. 4 of the DSM-RL requires reproductions made for the purposes of text and data mining. However, the reproductions in this case do not serve the purpose of further data analysis. Therefore, the copyright exceptions for text and data mining cannot be considered a basis for legitimacy. This limitation of purpose reflects the considerations underlying the DSM Directive. According to its recitals, the introduction of text and data mining limitations is intended not only to promote innovation and new technologies, but also to protect authors.
In addition to the infringements found in the GPT-4 and GPT-4o language models, the Munich Regional Court identified further infringements in the output. These infringements constitute an unauthorized communication to the public, among other things. Minor deviations from the original texts are irrelevant.
According to the Munich Regional Court, OpenAI directly commits the act of communicating to the public. OpenAI makes the work available to the public itself and does not merely act as an intermediary by providing infrastructure for third-party communication. Since OpenAI is responsible for the content of its models, which can be reproduced with simple prompts, the defendants cannot be compared to operators of hosting platforms, hardware providers, or software providers.
According to the Munich Regional Court, the outputs are not covered by the limitation provision for pastiche under Section 51a of the German Copyright Act. This provision requires exploitation for the purpose of pastiche. This entails artistic engagement with a pre-existing work or other reference object. However, this is not possible with the models due to their lack of personality, which they could not express artistically through the use of protectable elements.
The Munich Regional Court also affirms a claim for damages. In the opinion of the court, OpenAI acted negligently. Although the scope of the right of reproduction and the text and data mining limitations have not yet been fully clarified by the courts. However, in such circumstances, the negligence does not cease to exist on the basis of an innocent error of law. Such an error can only be considered if, when exercising the care required in business transactions, the person who made the error did not have to expect a different assessment by the courts. Normal litigation risk does not exonerate them. There is a legal duty to carefully examine the legal situation and seek legal advice when necessary.
The question of whether training large language models can rely on text and data mining exceptions under the DSM Directive was quite controversial (Marcus von Welser, A European View on A European View on Generative AI and Copyright, published in Landslide, Volume 17, Number 4, 2025, Pages 23-29, https://www.americanbar.org/content/dam/aba/publications/landslide/landslide-v017n04-2025.pdf ).
The Munich Regional Court’s decision thoroughly explains how large language models work. The decision also presents compelling reasons why the text and data mining exceptions do not apply, based on the wording of the relevant German and European legal provisions.
