AI-Training and German Copyright Law: OpenAI Loses

The Munich District Court I has redefined the boundaries of copyright law in the context of AI training with a landmark judgment (Case No. 42 O 14139/24) dated November 11, 2025. In the case of GEMA v. OpenAI, the court ruled that the memorization of protected works in language models (such as ChatGPT) constitutes copyright infringement under German Law – and that the text and data mining exception under Section 44b of the German Copyright Act (UrhG) does not cover AI-training.

The Case: GEMA v. OpenAI

GEMA (Gesellschaft für musikalische Aufführungs- und mechanische Vervielfältigungsrechte), a German collecting society, sued two companies of the OpenAI Group for injunction, disclosure, and damages, due to alleged copyright breaches.

The subject of the lawsuit is nine well-known German song lyrics. GEMA argued that the lyrics of these songs were memorized in the language models GPT-4 and GPT-4o, and that ChatGPT reproduced them verbatim in response to simple user queries (“prompts”). Both the memorization by the language models and the reproduction of the song lyrics in the chatbot’s outputs constituted infringements of exploitation rights.

“Memorization” describes the process where language models not only extract information from a training dataset during their training, but where, after training, a complete reproduction of the respective training data can be found in the model’s specified parameters.

In addition to copyright claims, GEMA also asserted violations of personality rights due to incorrect attribution of modified song lyrics. Personality rights (Persönlichkeitsrechte) are a cornerstone of German law, protecting an individual’s right to control their personal identity, including the right to be correctly identified as the author of a work. However, the court dismissed these claims – the judgment focuses exclusively on the copyright dimension.

OpenAI’s Defense Strategy

OpenAI argued that:

• The language models did not store or copy specific training data, but rather learned patterns from the entire training dataset reflected in the model’s parameters.
• Users, not OpenAI, were responsible for any infringing outputs, since the lyrics were only generated in response to user prompts.
• Moreover, any legal infringements were covered by copyright exceptions, particularly the text and data mining exception under Section 44b UrhG.

The court was not convinced by this line of argument.

The Ruling: Reproduction through Memorization is a Copyright Infringement under German Law

The court found that the song lyrics are reproducibly memorized in the language models GPT-4 and GPT-4o. This memorization constitutes reproduction under Section 16 UrhG and Article 2 of the EU Copyright Directive (“Directive on the harmonisation of certain aspects of copyright and related rights in the information society – InfoSoc Directive”). It is irrelevant that the texts exist only as probability values in the model parameters. This is sufficient for copyright-relevant reproduction.

Reproduction and Access to the Public Through Outputs

The court rejected OpenAI’s central line of defense, holding that OpenAI, not the users, was responsible for the infringement. OpenAI selects the training data, develops the model architecture, and controls the memorization process. User prompts were merely simple queries – it is the language models themselves that generate the concrete content of the outputs. Therefore, the models, not the users, are the determining cause of the copyright infringement.

Text & Data Mining Exception: Why It Doesn’t Apply in the Case of Memorization

Concerning the legally sensitive question of the applicability of the text and data mining exception (“TDM exception”) under Section 44b UrhG, the court initially acknowledged that language models generally fall within the scope of application of the TDM exceptions. However, if during the training of language models not only information is extracted from training data, but entire works are being reproduced, this does not constitute text and data mining within the meaning of the exception. The prerequisite of text and data mining and the respective exception provisions – that the automated analysis of mere information itself does not affect exploitation interests – does not apply in this constellation. On the contrary, the reproductions in the model constitute an infringement of the rights holders’ exploitation rights. The court left no doubt that memorization is something fundamentally different from the mere extraction of statistical patterns or information.

Analogy Due to Lack of Comparable Interest Situation?

The court also examined whether an analogous application of the TDM exception could be considered – and denied it. Even if one were to assume an unintended regulatory gap – because the legislator may not have been aware of the possibility of memorization – there would be no comparable interest situation. In the case of reproductions in the model, the exploitation of the work is substantially impaired, and the legitimate interests of rights holders are thereby violated. With an analogous application of the TDM exception, the authors and rights holders would be left without legal protection.

Consent of Rights Holders Through Publication?

The defendants’ interference with the plaintiff’s exploitation rights was also not justified by the consent of the rights holders. The court clarified that the training of AI models cannot be considered a customary and expected use that the rights holder must anticipate. The mere publication of content on the internet does not imply consent to its use for AI training purposes.

The Consequences: Injunction, Disclosure, and Damages

The Munich District Court largely granted the claims asserted by GEMA. OpenAI was ordered to cease further use, to provide information about the extent of previous use, and to pay damages. The decision is not yet final. OpenAI can appeal.

Conclusion and Outlook: A Precedent for the AI Industry?

The decision of the Munich District Court focuses on language model training and thus on the differentiation between permitted extraction of information and prohibited memorization.

It is likely that OpenAI will appeal the judgment, and the issue of AI and copyright will therefore continue through all German instances. If GEMA were to win in the final instance, this would establish that artists (even in the age of AI) retain control over their works, and that language models cannot simply use them without facing legal consequences.