On November 11, 2025, the Munich Regional Court I (Landgericht München I, case no. 42 O 14139/24) issued a landmark ruling in a case brought by the German music rights society GEMA against two companies in the OpenAI group. The court held that both the storage of copyrighted song lyrics within the parameters of an AI language model — known as “memorization” — and the subsequent reproduction of those lyrics in the model’s output (“regurgitation”) constitute violations of German copyright law. The judgment is not yet final; OpenAI has filed an appeal before the Munich Court of Appeal (OLG München, case no. 6 U 3662/25 e). Even so, it sets a new European standard for the legal assessment of generative AI systems.
The background of the case is straightforward: GEMA, Germany’s collecting society that administers the copyright of music authors and composers on their behalf, discovered that ChatGPT was reproducing nearly complete lyrics of well-known German songs in response to simple user prompts — including “Atemlos” by Helene Fischer and “Wie schön, dass du geboren bist” by Rolf Zuckowski. These texts had been so deeply embedded in the model’s parameters during AI training that targeted prompts could elicit near-verbatim reproductions. GEMA claimed this constituted a serious infringement of its members’ copyright and sought an injunction, disclosure, and damages from OpenAI.
The court was asked to assess two novel legal phenomena not explicitly addressed in the German Copyright Act (Urheberrechtsgesetz, UrhG): memorization and regurgitation. The Munich Regional Court I defined memorization as the storage of copyrighted content within the mathematical parameters of an AI language model in a manner that allows it to be retrieved through specific inputs. Regurgitation refers to the actual reproduction of that content in the model’s output — the moment when the chatbot presents a protected text to a user. The court found both acts to be independent infringements of copyright.
On the question of memorization, the court held that it constitutes reproduction within the meaning of Section 16 UrhG. Under German law, reproduction covers any physical fixation of a work that renders it perceivable, directly or indirectly, to human senses. The court ruled that it is irrelevant whether lyrics are stored as complete files or are reflected in the form of statistical probability values across the model’s weights. What matters is solely whether the model is capable of reproducing the protected text in a recognizable form. In doing so, the court articulated a forward-looking, technology-realistic concept of reproduction that encompasses implicit forms of storage.
OpenAI’s primary legal defense rested on the text and data mining exception under Section 44b UrhG, which was introduced into German law in 2021 following the transposition of EU Directive 2021/790. This exception allows for the automated analysis of copyrighted works without a license — but only where the use is exclusively for the purpose of analyzing large datasets and does not affect the rights holders’ exploitation interests. The Munich Regional Court rejected the application of this exception. The court reasoned that Section 44b UrhG privileges only preparatory analytical acts — such as the machine-reading and indexing of texts. The reproducible anchoring of entire works within a model’s parameters, and especially their subsequent output, go far beyond this purpose and substantially impinge on the rights holders’ economic interests.
OpenAI’s subsidiary argument based on Section 57 UrhG — the “incidental inclusion” exception — also failed. OpenAI contended that the song lyrics were merely incidental elements of a vast training dataset and thus covered by the exception. The court disagreed: a training dataset is not an independent copyrighted main work in its own right, and the individual song texts are not dispensable elements of the training process but deliberately incorporated content with independent economic value.
Particularly significant is the court’s analysis of liability attribution. The Munich Regional Court I made clear that OpenAI as the operator of the AI systems is itself responsible for the memorization and regurgitation of protected content. The selection of training data, the design of the model architecture, and the technical configuration of the system are within the provider’s sphere of responsibility — not the user’s who submits a prompt. For businesses and individuals using ChatGPT or similar systems, this is initially a welcome clarification: primary liability for AI output lies with the provider of the system, not the end user.
However, the court explicitly cautioned that users who adopt, publish, or commercially exploit AI-generated content without review take on their own copyright risks. Anyone who publishes an unedited AI output or integrates it into a product may be independently liable for any copyrighted passages contained therein. The judgment therefore underscores the need for careful human review of AI-generated content — especially in creative contexts such as texts, song lyrics, poetry, or screenplay excerpts.
GEMA had also based part of its claim on alleged violations of the song authors’ moral rights (Persönlichkeitsrechte). The Munich Regional Court dismissed this part of the claim, holding that the memorization and regurgitation of song lyrics does not attribute false works to the authors and does not otherwise violate their personal rights. This partial dismissal demonstrates that the ruling is nuanced and not a blanket condemnation of AI technology — the case concerns copyright law, not the personal dignity of the creators.
The GEMA ruling is far more than a dispute between a German collecting society and a US technology company. It is Europe’s first landmark judgment on the copyright assessment of AI training data and model outputs, and it has far-reaching consequences for all companies that develop or deploy generative AI. If the judgment is upheld on appeal, AI providers will be required either to license their training data or to ensure their models do not memorize and reproduce protected content. Both paths are technically and economically demanding and are likely to fundamentally reshape the licensing landscape for copyrighted works.
For companies integrating generative AI into their products or using it as an internal tool, the time to review AI usage practices is now. Key steps include: examining which AI systems you use and whether their providers have secured licensing agreements with rights holders; implementing internal processes to review AI-generated output for potentially protected passages; seeking legal advice before commercially publishing AI-generated content; and monitoring the appeal proceedings before the OLG München and parallel cases in other European jurisdictions that will further clarify the copyright framework for AI.
The Munich Regional Court’s judgment against OpenAI marks a turning point in Europe’s engagement with AI and copyright. It establishes that generative AI models may not freely memorize and reproduce copyrighted works, and that the text and data mining exception of Section 44b UrhG is not a blanket authorization for the training and operation of such models. The case is not yet concluded; the appeal before the OLG München will be closely watched and is expected to resolve further fundamental questions. What is already clear today is that the deployment of generative AI requires not only technical diligence but also careful attention to copyright law.
Do you have questions about AI and copyright law, or about using AI tools in a copyright-compliant manner within your business? The lawyers at HUFELD PartGmbB are here to provide comprehensive advice. Get in touch now.
Are you looking for a specific topic?
.png)
.png)
