Copyright Meets AI: Munich Regional Court I Rules in GEMA vs. OpenAI Case

How far can providers of artificial intelligence go when it comes to using copyright-protected content?
This question is at the heart of a landmark legal dispute between GEMA, the German performance rights organization, and OpenAI, the operator of the well-known chatbot ChatGPT. On November 11, 2025, the Munich Regional Court I (Case No. 42 O 14139/24) ruled in favor of GEMA. In doing so, the court applied a new legal standard for evaluating the training and output of AI models. The decision is not only a (temporary) victory for rights holders but could also have a lasting impact on the development and use of AI systems in Germany and Europe.
 

The Case

At the center of the dispute were song lyrics from nine well-known German artists, including works by Kristina Bach, Herbert Grönemeyer, and Rolf Zuckowski. GEMA, acting as a collective management organization for authors’ rights, accused OpenAI of integrating these lyrics into its language models without authorization. According to GEMA, the models were trained on the basis of the protected content and could reproduce the lyrics almost verbatim upon user request.

OpenAI, however, argued that its models do not store specific data but instead learn patterns and probabilities from large datasets, generating outputs based on these patterns. OpenAI also relied on the copyright exception for text and data mining (TDM, § 44b of the German Copyright Act, UrhG), which permits the reproduction of copyright-protected works for analytical purposes.
 

The Decision: Infringement of Reproduction Rights

While international cases in the U.S. are often argued under the "fair use" doctrine, and UK courts frequently reject jurisdiction over "foreign training," the Munich Regional Court I explicitly focused on the technical finding of "memorization" and the reproducibility of copyright-protected works within the model. The court held that the decisive factor is whether a work (e.g., a song lyric) is "recognizably" and reproducibly stored in the AI system—regardless of how the reproduction is technically achieved or whether it is actually output.

The court found that such storage and the reproduction of the lyrics in the form of chatbot responses to user queries constitute unauthorized reproduction of the protected works, thereby infringing the rights holders’ exploitation rights.

The court also ruled that this "memorization" is not covered by the TDM exception, as it goes beyond the mere analysis of data for training purposes. The TDM exception only permits reproductions necessary for compiling a dataset for AI training (e.g., reproducing a work by converting it into another digital format or storing it in working memory). In this case, however, the works were reproduced through "memorization", which constitutes an infringement of the author’s exploitation rights. Given the complexity and length of the song lyrics, the court also ruled out the possibility that the reproduction occurred purely by chance. The court thus rejected an innovation-friendly, analogous, or extended application of the TDM exception.
 

Responsibility for AI-Generated Output

The court’s statements on the responsibility for AI-generated outputs are particularly noteworthy. OpenAI argued that users, not the company itself, should be held responsible for the creation of outputs (and thus for the reproduction of protected song lyrics), as these outputs were based solely on user inputs (prompts). However, the court disagreed, holding that OpenAI, as the developer and operator of the language models, is responsible for the storage of training data, which significantly influences the generated outputs. Responsibility, therefore, lies with OpenAI. 
 

Implications for the AI Industry

For companies like OpenAI, the ruling presents a significant challenge. It clarifies that AI providers cannot freely use copyright-protected content to train their models without authorization. The court’s focus on "memorization" and reproducibility—rather than just the input data or the output—marks a shift in how such cases are evaluated.

This decision has far-reaching implications for AI providers and rights holders. AI developers will need to scrutinize even more carefully whether and how data is used to train their models. Additionally, the increased liability risk for AI providers may lead to a reassessment of the broad liability disclaimers and contractual indemnity clauses commonly included in terms and conditions for end users.

In the European context, this ruling could establish a legal perspective that diverges from the "fair use" principle in the U.S.: those who train AI models with protected works on a long-term basis will likely need to acquire licenses in the future—a requirement that could set a precedent across Europe. At the same time, a restrictive interpretation of copyright provisions could hinder Germany’s and Europe’s competitiveness in the development and use of AI applications, potentially giving other countries or regions the opportunity to establish more flexible and favorable frameworks for transformative AI development and use.

The ruling is not yet legally binding, and it is expected that this decision will not be the final word in the legal dispute between GEMA and OpenAI. It remains to be seen how the higher courts will rule on the matter.