GEMA wins against OpenAI in a landmark case before the Regional Court of Munich

Although the full reasons for the decision are not yet available, the court has published a detailed press release (in German). Below, we summarise the background to the proceedings and the court's key statements.

1. Background

GEMA is a collecting society for authors of musical works. It currently has over 95,000 members in Germany and represents the rights of over two million rights holders worldwide. In a press release dated 13 November 2024, GEMA announced that it had filed a lawsuit against OpenAI with the Regional Court of Munich.

The background to the lawsuit is GEMA’s allegation that OpenAI trained its well-known product, ChatGPT, using copyright-protected song lyrics from the repertoire of GEMA members, including Helene Fischer and Rolf Zuckowski. GEMA claims that this can easily be proven because ChatGPT can reproduce the lyrics in full without further internet access. GEMA also claims that OpenAI systematically used copyright-protected material to train its AI model, knowingly accepting that this constituted a violation of the law. However, GEMA considers the use of song lyrics is only permissible in return for appropriate remuneration.

2. The Regional Court of Munich’s decision

From a legal perspective, it was anticipated that, similar to the proceedings brought by Robert Kneschke against LAION e.V. before the Regional Court of Hamburg, the Regional Court of Munich would primarily address the application of the “text and data mining” exception (“TDM exception”) in Sec. 44b of the German Copyright Act (UrhG). These expectations were met, in contrast to the Hamburg proceedings where – somewhat surprisingly – the more specific Sec. 60d UrhG took centre stage.

Initially, the court identified two instances of infringement of the copyright reproduction rights of the song lyrics in question.

Firstly, the “memorization” of the lyrics constituted reproduction. “Memorisation” occurs when the language models not only extracts information from the training data set during training but also adopts it in its entirety. The court considered this to be proven, given that the length of the lyrics meant an identical reproduction in the output could not be mere coincidence. Thereby, the court rejected OpenAI’s argument that its language models did not store or copy specific training data but rather reflected what they had learned based on the entire training dataset in their parameters.

The court then addressed the widely debated question of whether the TDM exception in in Sec. 44b UrhG applies. Ultimately, the court ruled that it did not. While the reproduction required for compiling the training data was generally covered by the provision as such preparatory actions do not affect any exploitation interests, the situation was different if copyrighted works are reproduced during the training of the model (i.e. if information is extracted from the training data). Such reproductions interfered with the exploitation rights of the rights holders. Conversely, the premise of the TDM exception was that only information is extracted, with no reproduction taking place that interferes with the exploitation rights of the right holders concerned. The court rejected another “presumably technology- and innovation-friendly interpretation” of the TDM exception due to the “clear wording of the provision”, as reproductions in the model would have a lasting adverse effect on the exploitation of the work, thereby infringing the legitimate interests of rights holders.

Finally, reproducing the song lyrics as output also constituted reproduction of the copyrighted works. In reaching this unsurprising conclusion, the court rejected OpenAI’s defence that the output data was generated solely as a result of user input (prompts) and that OpenAI was therefore not responsible for it. According to the court, OpenAI is solely responsible for the models’ architecture, memorisation and, ultimately, outputs.

3. Comment and outlook

The ruling, which is not yet final, is certainly good news for rights holders and surprisingly adamant in the wording of the press release. Conversely, its implications are not limited to OpenAI, but should also prompt other providers of language models to take notice.

Similar to the Hamburg LAION case, the GEMA case can be considered a test case among the “AI lawsuits”. These cases are as challenging as they are exciting from a legal perspective, as solutions must be found for a number of unresolved legal issues, and the individual cases often have specific characteristics. While the LAION case seems to focus on the somewhat specific limitation of Sec. 60d UrhG, the Munich GEMA case is likely to yield more generalisable findings regarding Sec. 44b UrhG. One thing is certain: if the parties wish, both cases will reach the Federal Court of Justice and the European Court of Justice.

The next step in this process is the announcement of the Higher Regional Court of Hamburg’ judgment in the LAION case, which is scheduled for 10 December 2025. We will keep you updated on our blog!