Facts
GEMA, the German collecting society for music rights, sued two companies in the OpenAI group over the lyrics of nine songs by well-known German authors, including Kristina Bach and Rolf Zuckowski. GEMA argued that the lyrics had been used to train the language models behind ChatGPT and were reproduced largely verbatim in response to simple prompts. It sought an injunction, disclosure of information and damages. It also claimed that attributing altered lyrics to their authors infringed their personality rights.
What was the legal question in GEMA v OpenAI?
The court had to decide whether the lyrics had been memorised, and thus reproduced, within the models; whether any such reproduction was covered by the text and data mining (TDM) exception in § 44b of the German Copyright Act (UrhG); and whether responsibility for reproducing the lyrics in outputs and making them available lay with users or with the companies operating the models. It also considered whether attributing altered lyrics to their authors infringed their general personality rights.
What did the Munich Regional Court I (Landgericht München I), 42nd Civil Chamber decide?
The court largely upheld the claims. Comparing the lyrics with the outputs, it found that memorisation had occurred and held that embedding the works reproducibly in the models amounted to reproduction. The TDM exception covers preparatory copies made for analysis; where works are permanently reproduced in the model, the authors' exploitation interests are affected and the premise of the exception no longer holds. Reproducing the lyrics in outputs also infringed the rights of reproduction and making available. OpenAI, which selected the training data, designed the architecture and was responsible for the memorisation, was liable, rather than the users. Liability for damages was established; the personality-rights claim failed.
Why does GEMA v OpenAI matter?
The judgment is among the first in Europe to assess the contents of a language model directly under copyright law. The court accepted that memorisation can be proved through outputs. In the English case of Getty Images v Stability AI, by contrast, the court found that the model did not store or reproduce the works it was trained on. The Munich court's narrow reading of the TDM exception will remain central to the debate whatever the outcome on appeal.
Related stages
On 8 December 2025 OpenAI appealed to the Higher Regional Court of Munich (OLG München), case no. 6 U 3662/25 e. According to GEMA, on 31 July 2026 the Munich Regional Court also ruled in its favour in a separate action against the music-generation service Suno.
The same question elsewhere
May AI be trained on copyright works?
Building the dataset fell within the text and data mining exceptions.
Kneschke v LAION e.V.Memorising lyrics in the model was reproduction, outside the exception.
— this decisionUsing headnotes to build a competing product was not fair use (non-generative AI).
Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc.India · 2026Yes, at interim stageStoring news content to train the model was held prima facie to be fair dealing; an appeal is pending.
ANI Media Pvt. Ltd. v. Open AI OpCo LLC