GEMA vs. OpenAI: Landmark ruling on AI language models and copyright law

In its ruling of November 11, 2025 (Ref. 42 O 14139/24), the Regional Court of Munich I ruled that OpenAI infringes the copyrights of well-known German artists with its AI language models and chatbots based on them. The AI language model has memorized song lyrics by well-known German musicians, such as "Atemlos" by Helene Fischer and Kristina Bach or "Männer" by Herbert Grönemeyer, and the chatbot based on it can reproduce them largely faithfully. This affects the artists' exploitation interests, which is why OpenAI's use of the song lyrics without a license, as has been the case to date, is not permissible.

Storage of song lyrics

First, OpenAI argued that the technical characteristics of the language model did not constitute a copyright infringement. Training data such as song lyrics were not stored or copied by the language model. Rather, the language model only reflected and learned based on the entire training data set.

The Regional Court of Munich I did not accept this argument. It is known from information technology research that training data can be contained in language models and can be extracted as outputs (so-called memorization). This is the case when language models not only extract individual pieces of information from the training data, but also adopt the training data in its entirety. The song lyrics were found to be reproducible in the language models. Due to the extensive texts, this could not be considered a coincidence.

Chatbot as the producer of the output

OpenAI also argued that the outputs were not generated by the chatbot itself, but rather by the user who issued the command (known as a prompt) to display the texts. The court countered this argument by stating that the song lyrics could be extracted using simple prompts. The providers are responsible for the architecture and training of their language models.

Not covered by the text and data mining exception to copyright law

OpenAI also argued that the activities of the language model constituted text and data mining. This would mean that they fell under the exemption provision of Section 44b of the German Copyright Act (UrhG) and were therefore permitted. Whether language models fall under data mining in terms of their functionality is a legal question that remains unresolved and highly controversial. 

The court first determined that the reproduction of song lyrics when compiling the data corpus for training was covered by text and data mining. This (initial) reproduction was done for analysis purposes, which did not affect the exploitation interests of the author.

However, the training itself was not carried out solely for analysis purposes. And since not only individual pieces of information were extracted, but entire works were reproduced, the exploitation interests of the authors were also affected. The output generated by the chatbot with users was also not solely for analysis purposes. Here, too, the entire work was reproduced and the exploitation interests of the authors were affected. The court rejected a corresponding or supposedly more innovation-friendly interpretation of Section 44b UrhG.

Not an insignificant accessory

The court also did not consider the song lyrics to be an insignificant accessory (limitation under Section 57 of the German Copyright Act) to the rest of the training data.  This would require the existence of a copyright-protected main work. However, the training data as such did not enjoy copyright protection.

No consent

The court also ruled that the training of language models was not a normal and expected type of use that the rights holder had to anticipate. It therefore also rejected consent.

Consequences of the ruling

The plaintiffs praise the ruling as a milestone in protecting musicians and other creative artists from the unlicensed use of their works by artificial intelligence. In fact, it is conceivable that the ruling will have implications far beyond the music industry. However, it should be noted that the copyright infringements were only recognized because the protected song lyrics could be reproduced and accessed in their entirety. This is not the case every time AI is trained with copyrighted material. Furthermore, the ruling is not yet final. It is likely that the parties will appeal. It would certainly be helpful if the Federal Court of Justice (and the European Court of Justice) could ultimately comment on this issue.