GEMA vs. OpenAI: Munich Regional Court I issues landmark copyright decision
Neither the storage of copyright-protected content in AI language models nor its reproduction in AI output are permissible under copyright law.
On 11 November 2025, Munich Regional Court I issued a ruling (case no. 42 O 14139/24) on the matter of whether the storage of copyright-protected content in the parameters of an AI language model (memorisation) and the reproduction of such content in AI output (regurgitation) is permissible under copyright law. The Court ruled that it is not.
The claimant in the case is the collecting society GEMA (Gesellschaft für musikalische Aufführungs- und mechanische Vervielfältigungsrechte – Society for musical performing and mechanical reproduction rights), which is asserting claims for injunctive relief, information and damages against two companies of the OpenAI Group. The claimant argues that the AI language models operated by the defendants and the chatbots based on them memorise copyright-protected song lyrics and reproduce them in the output. The storage and reproduction of the song lyrics constitutes an infringement of copyright by way of reproduction, communication to the public and adaptation of the same. The claimant additionally argues that this violates the personal rights of the songwriters.
Munich Regional Court I upheld the action with regard to the copyright claims asserted by the claimant. However, it rejected any violation of the personal rights of the songwriters. The memorisation and regurgitation of the lyrics does not attribute third-party works to the authors, nor does it result in any other violations of personal rights:
Memorisation and regurgitation
Munich Regional Court I found it to be a proven fact that training data can be contained in AI language models and can be extracted again in the output. According to the Court, this is the case when the parameters of the AI language model to be trained not only extract information from the training data during AI training, but also do so when the entire content is factored into the weighting of the parameters. One reason for this may be the recurrence of the same content in a training data set.
Whether a protected work is contained in the AI language model can be determined by entering prompts. Such prompts also enabled the lyrics to be reproduced in the output in the present case. The Court emphasised that it is possible to reproduce the lyrics in the output even using very simple prompts.
Munich Regional Court I: Memorisation is a reproduction pursuant to section 16 German Copyright Act (UrhG)
Munich Regional Court I found that the memorisation of copyright-protected content in the AI language model infringes the rights holders' rights of reproduction (section 16 German Copyright Act (UrhG)). The term "reproduction" covers any material representation of a work that is capable of making the work directly or indirectly perceptible to the human senses in any way. While direct perception allows the work to be easily recognised by the human senses, indirect perception requires additional intermediate steps, such as the use of technical devices.
Memorisation fulfils the aforementioned requirements for reproduction. Munich Regional Court I emphasises that it is irrelevant that the content is only contained in the language models in the form of probability values and distributed across various parameters. It is sufficient that the AI language model is capable of reproducing the content in a recognisable form as song lyrics. Although the works can only be perceived indirectly (by entering certain prompts), this does not change the classification of memorisation as "reproduction".
Regurgitation is an adaptation, reproduction and communication to the public
Additionally, the Court classified the regurgitation, i.e. the reproduction of the copyright-protected content in the output, as an adaptation pursuant to section 23 German Copyright Act (UrhG), as a reproduction pursuant to section 16 German Copyright Act (UrhG) and as a communication to the public pursuant to section 19a German Copyright Act (UrhG).
Pursuant to section 23 German Copyright Act (UrhG), the author has the right to exploit their work in a modified form. Thus, appearances of a copyright-protected work that deviate from the original are also protected up to a certain limit. The Court ruled that the song lyrics were contained in the output in a recognisable form and that any additions made by the AI language model (hallucinations) did not render them unrecognisable.
Furthermore, the Court also affirmed that the song lyrics in the output of the defendant's AI language models constituted communication to the public. Here, the Court emphasised that the interactive transmission of the protected song lyrics was made possible by the defendant's chatbot. In principle, it can be used by anyone at any time, provided they have internet access and an internet-enabled device.
No justification for infringements of rights by AI language model
The infringements of the claimant's rights, which it exercises on behalf of the songwriters, as determined by the Court, could not be justified.
In copyright law, infringements of rights can generally be justified in two ways. Either the rights holder has given their consent to the use of the rights or there is a legal permission (limitation) that allows their use. Munich Regional Court I rejected both justifications.
The Court emphasised that, for example, the text and data mining limitation in section 44b German Copyright Act (UrhG) did not provide any justification for the acts of use. The limitation allows acts of reproduction in the run-up to automated processes for analysing large data sets. The limitation is based on the idea that no exploitation interests of rights holders are affected by such preparatory acts.
However, since the training of the AI language models involved transferring not only abstract information relating to the song lyrics into the parameters, but also the song lyrics themselves, this is not merely a preparatory act that leaves the exploitation interests of the rights holders unaffected. In the Court's opinion, the limitation therefore does not permit memorisation in the AI language model or regurgitation in the output.
Munich Regional Court I also rejected the rights holders' consent to memorisation or regurgitation. Rights holders do not have to regard the memorisation and regurgitation of their protected works by AI language models as normal and expected acts of use, especially since the defendants themselves admit that these processes are actually rare errors of the AI language models.
First ruling on acts of use that are subsequent to training
The ruling of Munich Regional Court I is not yet final and absolute. It can be assumed that it will be subject to comprehensive review in the appeal proceedings. It remains to be seen whether the ruling will be overturned or amended.
Alongside the decision of Hamburg Regional Court from 2024, this ruling is the second significant ruling in Germany to address copyright issues in connection with AI systems.
However, unlike the ruling of Hamburg Regional Court, which dealt with the question of the admissibility under copyright law of the reproduction of works in the context of the creation of a training data set, the ruling of Munich Regional Court I deals with acts of use under copyright law which are subsequent to training. These are the storage of works in the AI language model and their reproduction in the output.
It should be noted that, to date, the question of whether acts of use under copyright law that are directly associated with AI training as such are also covered has not been clearly clarified.
Careful examination of AI-generated content required
The ruling by Munich Regional Court I shows that caution is required from a copyright perspective when creating and using creative content with the help of generative AI systems.
Even the use of simple prompts can, in certain cases, lead to the reproduction of copyright-protected training content in the output. If such content is used, published and exploited without being checked, the user may be exposed to claims for injunctive relief or damages under copyright law.
To avoid copyright liability risks, content generated with generative AI must always be carefully checked and, if possible, comprehensively revised by a human being.
