An analysis of the judgment of the Munich I Regional Court, dated November 11, 2025 – 42 O 14139/24 (GEMA ./. OpenAI)
The legal assessment of generative AI systems is currently one of the central challenges for European copyright law. With its ruling of November 11, 2025, the Munich I Regional Court has set one of the first comprehensive legal benchmarks for the copyright classification of AI models. In the proceedings GEMA vs. OpenAI The chamber affirmed both reproduction and public making available of protected works through the memorization and playback of song lyrics in the GPT-4 and GPT-4o models. The decision's reasoning extends far beyond the individual case and has a significant signaling effect for the training and operation of generative AI in Europe.
The following contribution examines the court's essential lines of argument, classifies them dogmatically, and assesses their scope for the future interpretation of copyright law in the context of artificial intelligence.
Facts and procedural background
The defendants were two companies in the OpenAI group, which operate language models and chatbots based on them. The plaintiff, GEMA, acting as a collecting society, asserted claims for injunctive relief, disclosure, and damages due to the use of nine well-known song lyrics in the training and outputs of the models. In response to user requests, ChatGPT had reproduced these lyrics almost verbatim. OpenAI denied that a reproduction had occurred, arguing that language models do not store specific training data, but merely probabilities.
The Munich I Regional Court did not follow this reasoning and considered a reproducible storage of the works in the model to have been proven. It confirmed both unlawful reproduction (§ 16 UrhG; Art. 2 InfoSoc Directive) and making available to the public (§ 19a UrhG).
What does „memorization“ mean in the context of copyright law?
1. Reproduction element for statistical models
The core of the decision is the assumption that training an AI model constitutes „reproduction“ if the training data is reproducibly contained within the model. The court bases this on the Union law concept of reproduction, which according to established ECJ case law is to be interpreted in a technology-neutral and extremely broad manner. Even indirect perceivability through technical means is sufficient.
With this, the court breaks with the reasoning of many AI providers who portray training as a purely mathematical abstraction. In the chamber's view, the relevant factor is that the lyrics were fully extractable from the model, not whether this describes the typical functionality of a model.
2. Differentiation from mere information retrieval
The court explicitly distinguishes between:
- extracted information (admissible), and
- adopted works impermissible.
This differentiation builds on the premises of text and data mining: a restriction regulation assumes that the work character disappears in the data processing operation. However, where – as here – the model is capable of reproducing the work almost 1:1, there is no longer merely „analytical“ use, but independent storage.
No application of the TDM limitation (§ 44b Copyright Act)
The court expressly rejects a justification beyond the limits of text and data mining. The regulations are tailored to preparatory analysis steps, not to permanent model content.
1. Purpose of the barrier
§ 44b of the German Copyright Act serves:
- the promotion of digital research,
- enabling automated evaluation of large datasets,
- while simultaneously protecting commercial interests.
The legislator expressly assumed that the work as such not is to be diversified.
2. No analog application
The chamber refers to the clear wording: The restriction only allows necessary technical intermediate steps. Permanent storage of complete works is not covered. The court also denies an unintended regulatory gap: the balance of interests is not comparable, as memorization directly impairs exploitation interests.
AI Provider Accountability
A special feature of the ruling is the clear assignment of responsibility to the operators of the models.
Users are not „manufacturers“ of the outputs
OpenAI argued that users are responsible for the content because it is generated by their prompts. The court does not agree: The models significantly shape the output. Users do not request specific texts, but merely trigger a result from the stored structures.
With this, the ruling addresses a widespread misunderstanding:
The AI provider remains responsible for the system it created itself, even if the user initiates the output.
2. The significance of training as a deliberate act of utilization
The defendants selected the training data and thus the works. The training of the models is an active process of use that cannot be attributed to user behavior.
Making publicly accessible through outputs
In addition to reproduction, the court also sees a public making available (§ 19a UrhG) because the AI makes the protected song lyrics retrievable for the general public. This constitutes an independent infringement and is also not covered by exceptions.
Evaluation and Importance for Future AI Regulation
The judgment is dogmatically far-reaching.
1. Significance for Generative AI
The ruling suggests that AI providers will no longer be able to argue in the future that models do not contain works, but merely parameters. The benchmark is: What is reproducible is duplicated.
This creates a clear boundary for the first time, which applies to all types of works – from music and texts to images and films.
2. Impact on the AI Industry
Upon confirmation of case law, licensing of large training corpora will become necessary. This would shift the balance of power between technology companies and the creative industries. AI providers would have to document comprehensibly:
- which data was trained,
- to what extent memorization takes place,
- what rights exist for this.
3. Relationship with the European AI Regulatory Framework
The collision with the AI Act is exciting. While it does not regulate copyright, it does contain transparency obligations that could also apply to training sets in the future. The ruling shows that AI regulation only works if copyright and AI regulation are considered together.
Conclusion
The decision of the LG Munich I is a milestone in the development of European AI law. It establishes fixed legal categories for dealing with generative AI and copyrighted works for the first time. Above all, the recognition of memorization as reproduction will decisively influence future models, training practices, and business models.
The decision also makes it clear that technical innovation does not take place outside of legal structures. When AI models act as creative tools, they must respect the same rules as any other form of using third-party content. The ruling thus creates an important balance between technological development and the protection of creative achievements – and could become the starting point for new, differentiated jurisprudence in the age of generative AI.