I. Background: AI Training and Output Using Song Lyrics
The Regional Court of Munich I has issued a landmark decision in the case GEMA v. OpenAI (42 O 14139/24), ruling that the training of large language models (LLMs) with copyrighted works and the reproduction of these works in the AI output constitutes copyright infringements. The 42nd Civil Chamber did not consider it necessary to refer the matter to the European Court of Justice.
The collecting society GEMA had sued OpenAI, claiming that ChatGPT was trained on copyrighted German song lyrics and that these lyrics appeared — sometimes verbatim — in the output of ChatGPT. The song lyrics included, among others, „Atemlos“ by Kristina Bach and „Wie schön, dass du geboren bist“ by Rolf Zuckowski.
II. Court’s Decision: Copyright Infringement Through Training and Output
The chamber upheld GEMA’s claims for injunction, disclosure, and damages. The court found that the lyrics used for training of the LLM were reproducible in OpenAI’s models GPT-4 and GPT-4o (so called memorization).
According to the judges, this memorization constitutes a reproduction under Article 2 of the InfoSoc Directive and § 16 of the German Copyright Act (UrhG), as the copyrighted material is fixed within the model’s parameters and can be extracted through outputs. The court rejected OpenAI’s defence that the models only reflect learned patterns rather than storing data.
Additionally, the reproduction and public communication of the lyrics in ChatGPT outputs were held to infringe exploitation rights by constituting a reproduction and a public disclosure. Responsibility for these infringements rests with the provider, not the users, since OpenAI designs, trains, and deploys the models that generate the outputs.
III. No Justification of Training Through Text and Data Mining (TDM) Exception
The court dismissed OpenAI’s reliance on TDM exception under § 44b UrhG and Article 3 DSM Directive with regard to the training of the LLM. These provisions permit reproductions only for analytical purposes and do not extend to memorization or reproduction of entire works.
The Chamber justified this with an interpretation in line with the DSM Directive and with the wording of the TDM exception. The TDM exception shall not undermine the economic interests of rights holders by leading to an exploitation of their works. The judges also ruled out an analog application of the TDM exception, stressing that AI model training involving memorization goes beyond what the legislature intended to permit.
IV. Significance and Outlook
While the Regional Court of Hamburg’s Laion ruling dealt with the collection of training data (310 O 227/23 under appeal. Decision by the Hamburg Court of Appeal is scheduled for 10.12.2025). The Regional Court of Munich’s ruling specifically addresses the use of copyright protected works for the training process and confirms that memorization constitutes a reproduction and moreover that output reassembling training data constitutes an infringing act of reproduction and public communication of that content. According to last weeks Getty Images v Stability AI ruling, Getty seems to have been unable to establish that Stability AI created clearly memorized (nearly identical) output. In any event Getty Images withdrew its training infringement claim, so that the ruling remains silent on that point.
The Regional Court of Munich’s ruling establishes that AI providers must ensure lawful use of copyrighted materials and can in particular not rely on broad readings of the TDM exception.
While the decision is not yet final and OpenAI may appeal, it sends a strong signal in favor of rights holders. When announcing its ruling, the chamber made clear that it is aware of the significance of the decision for the AI industry.