05 February 2026

The Show is Over (for the Time Being): OpenAI Needs a Licence for Song Lyrics

  • Articles
  • Compliance
  • Legal
  • Data / Technology / IP

The Munich I Regional Court rules: Artificial intelligence may not reproduce copyrighted song lyrics without a license. Operators are liable when full lyrics are generated in response to simple user prompts.

  • Luca Hitz

    Legal Partner

  • Anna Schweizer

    Legal Associate

In the dispute between the German collecting society for musical performance and mechanical reproduction rights (GEMA) and the US software company OpenAI, the Munich I Regional Court has drawn a clear line: even artificial intelligence may not use copyright-protected song lyrics without permission. Anyone reproducing entire song lyrics and making them available to chatbot users requires a licence – especially if this happends in response to simple queries («prompts»).

OpenAI argued that its language models did not store or copy specific training data, but merely reproduced what they had «learned» during training. The company claimed that the users, not OpenAI, were responsible for any displayed content. Moreover, it claimed that any infringements were covered by copyright exceptions, in particular the exception for text and data mining (TDM). The court rejected this argument.

Content of the Ruling

According to the court, OpenAI violated copyright law by using copyright-protected song lyrics without prior authorisation both for training and operating its GPT-4 and GPT-4o models. The decisive factor was not the training itself, but the models’ ability to reproduce entire song lyrics almost word for word in response to simple user prompts.

The judges based their decision on the following considerations:

  • Memorisation instead of mere analysis: GEMA successfully demonstrated that training data is not only processed statistically, but embedded in the model in a way that amounts to permanent storage. This «memorization» is legally relevant, as it effectively enables reproduction.
  • Responsibility of the operator: As the operator and architect of the models, OpenAI bears responsibility when users can easily retrieve works in which the creative elements of the original song lyrics are clearly recognisable in the model’s outputs.
  • Limits of the TDM exception: The invocation to the text and data mining exception was unsuccessful. According to the court, the reproductions contained in the model go beyond the purpose of pure information gathering. The storage of protected works within the model infringes the copyright holders’ exploitation rights, as it sustainably their economic exploitation of the works.
  • No violation of personal rights: While GEMA successfully enforced its copyright claims, the court dismissed the lawsuit regarding alleged violations of the artists' personal rights due to the misattribution of distorted lyrics. Text alterations caused by so-called «AI hallucinations» did not constitute a serious infringement of an individual’s social sphere (the professional and political activities of the individual).

Significance Beyond the Individual Case

This judgment marks a turning point. It is one of the first European landmark rulings on generative AI and makes clear that learning systems do not operate in a legal vacuum. The often-cited «black box» loses its aura once as it reproduces works owned and created by others for which the model operator has not acquired any rights.

Outlook

In early December 2025, OpenAI filed an appeal. The next instance is the Munich Higher Regional Court. The case could eventually reach the Court of Justice of the European Union, which would then decide whether and to what extent the memorisation of protected works constitutes a copyright infringement. Meanwhile, GEMA is preparing additional lawsuits – including against providers of music-generating AI systems.

 

 

 

Click here to learn more about our expertise:
Back