ChatGPT output of song lyrics infringes copyright: Munich Court

The Munich Regional Court I (Landgericht München I) recently held that ChatGPT infringed copyright by allowing the memorisation of song lyrics within its large language models (LLMs) and subsequently reproducing those works in the chatbot's outputs. (Gema v. OpenAI)

The decision addresses liability concerning both the training processes and the generated outputs of generative Artificial Intelligence (AI). The court largely ruled in favour of GEMA, a music rights organisation, concerning nine German song lyrics.

The judgment was handed down by the 42nd Civil Chamber, presided over by Judge Elke Schwager.

The lawsuit, filed against two entities of the OpenAI group (one based in Ireland, the other in the USA), centred on the unauthorised use of nine German song lyrics for which GEMA holds the exclusive usage rights. The texts included well-known works by artists such as Herbert Grönemeyer, Kristina Bach and Rolf Zuckowski.

GEMA’s claim focused on two distinct acts of infringement. First, the reproduction of the lyrics in the LLM during the training phase, which resulted in the "memorisation" of the protected texts within the models.

Second, the subsequent public communication and reproduction that occurred when the ChatGPT chatbot generated the lyrics, in full or in part, in response to simple user prompts (e.g., "How long is the text of [Song Title]?").

The defendants argued that the models do not store or copy specific data, but instead learn statistical patterns. Therefore, no fixed copy as required by copyright law exists. They further asserted that the training process was justified under the Text and Data Mining (TDM) exception. Finally, they contended that liability for any infringing output should fall on the end user who provided the prompt, not the platform operator.

The court, however, noted that the AI models are trained on vast datasets and the issue arose because the models could "regurgitate" the copyrighted works, often almost verbatim, upon simple prompting. The plaintiff’s process servers successfully demonstrated this by generating the lyrics using easily formulated prompts.

The court largely upheld the infringement claims based on the following rationale:

Reproduction within the model (memorisation)

• The court ruled that the memorisation of the lyrics within the LLM parameters constitutes a reproduction because the lyrics are "reproducibly contained". The fact that the fixation occurs as numerical probability values was deemed irrelevant for copyright purposes.

• The TDM exception was deemed inapplicable to this. The court reasoned that TDM permits copies only for the purpose of analysis to extract information. The long-term memorisation and reproduction of a complete or substantial work exceeds this purpose, infringing upon the creator's exploitation rights.

Public communication via outputs

• The court found the generation of lyrics in the chatbot’s outputs to be an unauthorised act of making the work publicly available a form of public communication.

• The AI provider was held directly liable; the company, not the user, was said to have committed the act of making the work publicly available. The court stated that simple prompts do not transfer liability for the infringing content to the user.

OpenAI was found to be liable for damages to the tune of 4,620.70 Euro. It was ordered to cease and desist from the infringing activities, and provide GEMA with disclosure on the extent of the infringement

The court found that the defendants acted with at least negligence, noting that they were aware of the risk of memorisation in their models since at least 2021.

Due to the finding of negligence, the court rejected the defendants' requests for a six-month grace period for compliance and dismissed arguments that the claim was disproportionate.

It dismissed GEMA's ancillary claim regarding the violation of the authors' general personal rights concerning the attribution of altered lyrics.

AI translation disclaimer: This news article is based on a translation of a German legal judgment. The original text was processed and translated using an Artificial Intelligence (AI) system. While effort was made to accurately convey the legal content and terminology, discrepancies may exist between this version and the original German text, which remains the sole official and legally binding document.

[Read Judgment]