The judgment delivered by the Regional Court of Munich I on 11 November 2025 (Case No. 42 O 14139/24, hereinafter “the Judgment”) in the litigation between the German collective management organisation for musical performing and mechanical reproduction rights (“GEMA”) and the U.S. software company OpenAI, Inc. (“OpenAI”) has attracted significant public attention. Not only the legal community but also the general press and the broader public have taken a considerable interest in the case. This is due to the particular significance of the ruling for creative and cultural professionals and for Europe as a business location, especially for international technology companies.
It is the second decision by a German court addressing the copyright implications of generative artificial intelligence (“AI”) and formulating initial guiding principles for assessing the permissibility of using copyright-protected data in AI models and systems.
The first “AI judgment,” issued by the Regional Court of Hamburg on 27 September 2024 (Case No. 310 O 227/23), concerned the creation of AI training datasets (first phase) and therefore addressed the “preparatory stage” of AI use. In that judgment, the court found that the reproductions associated with dataset creation could, in principle, be permissible under §§ 60d, 44b German Copyright Act (hereinafter “UrhG”) —the statutory limitations for text and data mining (“TDM”). That judgment is not final; the plaintiff has filed an appeal with the Higher Regional Court of Hamburg.
The recent judgment of the Regional Court of Munich I now addresses the actual training of the neural network using the dataset (second phase) and the subsequent use of the trained AI system (third phase). The court granted GEMA’s claims for injunctive relief, information, and damages, finding that OpenAI had unlawfully exploited nine well-known German song lyrics (including “Atemlos” by Kristina Bach and “Wie schön, dass du geboren bist” by Rolf Zuckowski) both in the KI model itself and through the reproduction of the lyrics in the chatbot’s output.
The court’s reasoning begins with the conclusion of the 42nd Civil Chamber that the song lyrics at issue are reproducibly contained in OpenAI’s language models. According to the court, this means that the model does not merely extract and process metadata or statistical parameters but instead memorises the training data— including the protected lyrics— in its entirety. In computer science research, this phenomenon is known as memorization (cf. para. 169 et seq. of the Judgment).
The court inferred such memorization by comparing the song lyrics included in the training data with the model’s output. Because the outputs matched and given the complexity and length of the texts, random coincidence was ruled out. The court therefore assumed an intentional reproduction of copyright-protected content.
Based on this, the court articulated an important presumption for the copyright assessment of generative AI systems: If an AI model reproduces complex copyright-protected content in a consistent and reproducible manner, and no plausible random cause exists, it may be presumed that the model has memorised the work (para. 176 of the Judgment).
This presumption may help overcome the typical evidentiary challenges associated with generative AI—particularly regarding whether, which, and in what form (mere metadata or actual stored content) training data has been stored within an AI model. The court grounds this presumption both in findings from computer science research and in the readily verifiable model output, which it treats as a decisive indicator of the reproducibility of the protected content.
On the basis of the established memorization—amounting to a fixation of the copyright-protected linguistic works—the court held that the training process involves an act of reproduction within the meaning of § 16 UrhG (cf. paras. 176 et seq. of the Judgment).
The court emphasised that the fact that the model represents the data only through probability parameters is irrelevant (para. 182 et seq.). New technologies such as language models are covered by the right of reproduction under Art. 2 of the InfoSoc Directive and § 16 UrhG. According to the case law of the Court of Justice of the European Union (CJEU), even indirect perceptibility suffices for a reproduction if the work can be perceived with the aid of technical devices.
The court further held that this reproduction is not covered by the TDM limitations under § 44b UrhG nor by § 57 UrhG (insignificant accessory).
While AI models may in principle fall within the scope of the TDM limitation if the reproduction serves solely analytical purposes and does not interfere with the rightholders’ legitimate exploitation interests (para. 204 et seq.), the court found that this condition was not met. OpenAI’s models do not merely extract information; rather, they reproduce entire works. The output therefore competes directly with the original works—here, the lyrics— and significantly impairs their exploitation. The rightholders’ legitimate interests are thus infringed, and § 44b UrhG is not applicable.
A privileged use under § 57 UrhG was likewise excluded, as no principal work existed (para. 214 f.).
In sum, the Regional Court of Munich I concluded that the right of reproduction under § 16 UrhG had been infringed.
Beyond the training process, the court also examined whether the chatbot’s output constitutes a copyright infringement (third phase). This question was answered in the affirmative. Because the disputed song lyrics were recognisable in the outputs, the court held that both the right of reproduction under § 16 UrhG and the right of making available to the public under § 19a UrhG were infringed (para. 230 et seq.).
For § 19a UrhG to apply, it suffices that the work is made available in such a way that members of the public may access it from a place and at a time individually chosen by them, regardless of whether they actually access it (para. 259 et seq.). This requirement was fulfilled, as OpenAI, as operator of both the model and the chatbot, enables users to access the memorised lyrics. Such mediated communication also qualifies as public because the chatbot is accessible to anyone with internet access and an internet-enabled device.
The court also affirmed OpenAI’s direct liability as the operator and perpetrator of the infringing acts (para. 275). OpenAI selected the training data, designed the model architecture, and was responsible for memorization of the training data. This resulted in a decisive influence over the generated outputs. Thus, the court rejected OpenAI’s argument that only the user issuing the prompt was responsible for generating the output.
Finally, the court held that none of the statutory limitations under §§ 44b, 51, 51a, 53 (1) UrhG applied to the outputs.
The Judgment is not yet final. Given the importance of “artificial intelligence” and the conflicting interests of rightholders and AI model operators, it is expected that the numerous legal questions arising under §§ 16, 19a, 44b and 57 UrhG—as well as the underlying functioning of AI models—will be examined in further instances, possibly up to the CJEU.
Nevertheless, the Judgment already provides important guidance that may help reduce existing legal uncertainty. The court strengthens the rights of authors and rightholders while recognising the technical characteristics of neural networks, which do not necessarily memorise entire works. The reasoning on the inapplicability of § 44b UrhG aligns with the requirements for the TDM limitation as set out in the obiter dictum of the Regional Court of Hamburg (paras. 65 et seq. of the decision of 27 September 2025). That court also confirmed that text and data mining for analytical purposes remains permissible within the limits of § 44b UrhG.
Thus, Europe continues to offer scope for the training and further development of AI and software technologies. However, the decisive boundary is crossed where training datasets are incorporated in such a manner that the model output competes directly with the original work, thereby infringing the legitimate interests of the rightholders.
It remains to be seen how German and European courts will address these issues in future. A continuation of the case law of the Regional Court of Munich I can be expected in 2026— notably through the pending parallel case in which GEMA has brought an action against Suno, Inc.
