On 11 November 2025 the 42nd Civil Chamber of the Munich Regional Court I ruled in the case brought by GEMA, the German music copyright collective, against OpenAI. The court held that the model’s “memorisation” of lyrics and the chatbot’s output of lyrics constituted copyright infringement.
Two layers of the ruling
First, memorisation at the training stage is itself infringing. The court did not stay at the abstract question of whether training amounts to fair use; it focused on the model’s “memorisation” of protected works. Where the model can reproduce protected content, that act of memorisation already falls within the reproduction right.
Second, reproducing lyrics at the output stage is also infringing. The court separately held that the chatbot’s output of lyrics in its answers constitutes reproduction and making available to the public, independent of the memorisation at training.
Treating training and output as two independent infringing stages is the most methodologically valuable part of the ruling. It means a company cannot cut its liability link to training by arguing that “the output was triggered by the user”.
Why the ruling matters
European AI copyright discussion had largely stayed at the legislative and policy level, lacking binding judicial determinations. The Munich ruling provides a concrete line of reasoning on “whether model memorisation of protected content infringes”, and is likely to be cited in later cases.
It also resonates with the transparency duties under the EU AI Act: providers of general-purpose AI models must disclose a summary of training data, enabling rights holders to assert their rights. See our note on the EU AI Act transparency rules taking effect.
Four points for training-data compliance
First, training data provenance must be traceable. Since the ruling builds infringement on “memorisation”, a company must have a documented record of corpus sources and authorisation chains.
Second, de-duplication and filtering are not optional. Where the corpus contains large amounts of protected works, the probability of model “memorisation” rises markedly. De-duplication, filtering and copyright content identification at the cleaning stage are substantive risk-reduction measures.
Third, output-side interception is needed. For requests clearly aimed at full reproduction of protected works, the product should be able to identify and intercept, and to retain handling records.
Fourth, allocate ownership and liability in contracts. When procuring third-party models or corpora, the contract should specify liability allocation and indemnity. See our IP protection service.
Contrast with Chinese rulings
Chinese courts are also accumulating experience on generative AI copyright boundaries. The first AI model infringement case and the Shanghai AI prompt copyright case explored the model-output and prompt-creation angles respectively. See our notes on the first AI model infringement case and the AI prompt copyright case.
The paths differ, but the common trend is that AI copyright liability is not automatically exempted by “technological neutrality”. Companies need to move copyright compliance from after-the-fact legal response to the data and product design stage.
This article is based on public reporting of the Munich Regional Court I ruling in GEMA v OpenAI. It is for general reference only and does not constitute legal advice.