Judgment of November 11, 2025 – 42 O 14139/24 (nrkr)
In its judgment of November 11, 2025, the Munich Regional Court I essentially upheld the claims asserted by the collecting society GEMA against two companies of the Open AI Group. The decision marks a milestone in Europe for the copyright treatment of large language models (LLMs) and is likely to have a significant impact on training practices in the AI industry.
Core of the legal dispute
GEMA had determined that song lyrics by several German authors – including “Atemlos,” “Über den Wolken,” “Bochum,” and “In der Weihnachtsbäckerei” – were included in OpenAI’s training data. The lyrics could then be retrieved “largely true to the original” by ChatGPT through simple user queries. GEMA considered this to be unauthorized reproduction and public disclosure of the works.
OpenAI denied storing specific training data and pointed out that outputs were only generated on the basis of user prompts. Any interventions were also covered by the restrictions for text and data mining (TDM).
Court decision
The Regional Court of Munich I did not follow these arguments and clarified:
Memorization in AI models is a reproduction within the meaning of Art. 2 InfoSoc-RL and § 16 UrhG
The 42nd Civil Chamber is convinced that the song lyrics at issue are reproducible in the ChatGPT-4 and 4o models. Memorization is evidenced by a comparison of training texts and generated outputs. Given the complexity and length of the texts, mere coincidence can be ruled out.
This means that the works are embodied in the model parameters. The judge emphasized that new technologies such as AI models are also covered by reproduction rights. Even indirect perceptibility via technical aids is sufficient.
Text and data mining restrictions do not apply
The court clearly rejected a broad interpretation of the TDM restrictions (Section 44b UrhG):
- The limitation only allows preparatory analysis activities that are not relevant to the work.
- However, OpenAI’s AI training not only extracted information, but also reproduced the works themselves.
- This constituted a direct infringement of exploitation rights.
- An analogous application of the restriction was not possible due to a lack of comparable interests.
The decision emphasizes that memorization is relevant to copyright law and may not be carried out without remuneration.
Unauthorized public disclosure in the outputs
The output of texts by ChatGPT also constitutes reproduction and making available to the public, which is relevant to copyright law.
Important: The chamber considered OpenAI itself—and not the users—to be responsible. User prompts such as “What is the text of ‘Bochum’?” are not sufficient to attribute the infringement to the user.
No protection under Section 57 UrhG (insignificant accessory)
The song lyrics are not “incidental” because there is no “main work” within the meaning of the provision. The entire training data set is not a copyright-protected work.
No infringement of general personal rights
The court rejected only one point: GEMA was unable to enforce incorrect attributions or falsified song lyrics as a violation of personality rights.
Clear statement by the presiding judge
It is noteworthy that the judge clarified that training data in AI models generally remains permanent unless the entire data set is deleted. The requested transition period for continued use was therefore rejected.
Assessment and significance for practice
The ruling is not final; OpenAI has announced that it will appeal. Nevertheless, this is the first comprehensive European decision on the copyright admissibility of AI training with protected works. The signal effect is considerable:
- AI companies must expect that training with copyrighted texts without permission is illegal.
- The decision strengthens the position of collecting societies and rights holders in upcoming license negotiations.
- It significantly limits the scope of TDM barriers, especially in the commercial AI context.
- In future, AI providers must actively avoid memorization risks in training processes or develop other legally compliant solutions (e.g., licensing, filter mechanisms, editorial obligations).
Outlook
Whether and to what extent higher courts confirm the decision will be of strategic importance for the European AI market. In view of parallel regulatory efforts at the EU level – in particular through the AI Act – pressure on AI providers to introduce transparent and legally compliant training processes is likely to increase further.
Specific instructions for dealing with AI systems following the ruling of the Regional Court of Munich I
Companies that develop, train, or integrate AI systems should note the following points:
Audit training data
- Clarify which data sources are used for training, fine-tuning, or RAG processes.
- Avoid using texts or other works for which no rights or licenses are available.
- Document data provenance and rights chain.
Actively manage memorization risks
- Keep in mind that models can memorize and reproduce protected content.
- Use technical methods to “de-memorize” or minimize reproducibility (filters, RLHF masking, retrieval approaches).
Only train models with legally compliant data sets
- Give preference to licensed, public domain, or contractually clarified data pools.
- For open-source models: Check the terms of use and possible copyright restrictions.
Control AI outputs within the company
- Check whether the systems used reproduce protected content.
- Implement internal usage guidelines (“AI Usage Policies”), especially for marketing, product texts, and customer communication.
Adapt contract design
- Development, license, or integration contracts should clearly regulate:
- Data rights and sources,
- Distribution of liability in the event of copyright infringements,
- Assurances regarding memorization prevention,
- Obligations to delete or renew models in the event of legal violations.
Include risks arising from AI regulation (EU AI Act)
- AI providers must ensure detailed documentation, transparency disclosures, and copyright compliance in the future.
- Violations can result in heavy fines – in addition to civil law claims.
