Summary
In the case of GEMA v OpenAI entities operating ChatGPT, the Regional Court of Munich I held that the operators of a generative AI chatbot can be liable for copyright infringement when the system reproduces song lyrics in response to simple prompts. The court found that the chatbot produced outputs containing protected lyrics from songs represented by the collecting society GEMA. It held that such outputs may amount to reproduction of copyrighted works and that the operators can be liable because they exercise control over the functioning of the system that generates the outputs.
Copyright Infringement Through AI Outputs
The plaintiff, GEMA, is a German collecting society that manages the rights of composers, lyricists, and music publishers with respect to musical works and song lyrics.
The defendants are companies belonging to the OpenAI corporate group, which develop, operate, and license generative artificial intelligence technologies including AI language models and chatbot systems that generate text responses to user prompts.
GEMA alleged that the chatbot reproduced song lyrics belonging to its repertoire when users entered prompts requesting the lyrics of specific songs. The plaintiff also alleged that the disputed lyrics had been included in the training data used to train the defendants language models.
Questions Before the Court
1. Whether the outputs generated by the AI chatbot that reproduce song lyrics constitute reproduction of copyrighted works.
2. Whether the operators of the AI language model can be held liable for copyright infringement arising from such outputs.
3. Whether comparison between the original works and the chatbot outputs can establish memorization of the works within the AI model.
Arguments Presented By the Parties
Plaintiff Arguments
The plaintiff argued that the chatbot reproduced copyrighted song lyrics in its outputs when users entered prompts asking for the lyrics of specific songs. According to the plaintiff, these outputs showed that the lyrics had been memorized in the model during training.
The plaintiff further argued that the reproduction of lyrics through chatbot outputs created copies both on the user device and in the chat history stored on the defendants servers.
The plaintiff also argued that the defendants exercised significant control over the AI system. According to the plaintiff, the defendants determined the training data, designed the training process, and configured the decoding architecture that produced the outputs.
Defendants Arguments
The defendants argued that the outputs generated by the chatbot are probabilistic sequences produced through statistical processes within the language model.
They stated that the model itself produces probability distributions rather than final text and that the displayed output results from decoding mechanisms that introduce randomness and variation.
The defendants also argued that the model parameters remain fixed after training and that the outputs depend on user prompts and sampling mechanisms such as temperature settings and token selection.
Court Analysis of AI Generated Outputs
The court stated that copyright infringement may arise when a chatbot reproduces protected works in its outputs. According to the court, when a user enters a simple prompt requesting the lyrics of a song and the chatbot produces the lyrics in response, such output may amount to reproduction of the protected work.
The court observed that comparing the original work with the output generated by the chatbot allows the court to determine whether the model has memorized the work. According to the court, such comparison can establish whether the output reflects stored content derived from the training data.
The court further stated that operators of AI language models may be liable for copyright infringement resulting from chatbot outputs when they exercise control over the system. The court observed that although users provide prompts, the operators design the training process, select the training data, and determine the architecture and mechanisms that produce the outputs.
The court also stated that where the outputs arise from simple prompts rather than complex manipulation by users, the operators retain sufficient control over the act that results in the reproduction of the protected work.
Findings
The court ordered the defendants to refrain from reproducing or making available the disputed song lyrics through the outputs of the chatbot.
The court also ordered the defendants to provide information regarding the extent of the infringing acts and the revenue generated from them.
The court further held that the defendants are liable to compensate the plaintiff for damages arising from the infringing acts.
Relevant Paras
Para 168 to 173
By comparing the original work with the output of a simple prompt, the court can determine to the satisfaction of the court pursuant to Section 286 of the German Code of Civil Procedure whether the work in question has been memorized.Para 181
The memorization of linguistic works in an AI language model constitutes reproduction under Section 16 of the German Copyright Act because the work is physically fixed and can be made indirectly perceptible.Para 184 to 185
For a physical fixation to exist, it is not necessary that a clearly definable data set be identified in the model. Even if the work is decomposed into parameters, a physical fixation exists if the parameters are found in the model.Para 193
Reproduction during the creation of the training data material falls under the limitation of Section 44b of the German Copyright Act but not during the training of the model because this does not solely serve to prepare for text and data mining.Para 275 to 277
The operators of the language model are liable for copyright infringement through outputs because they exercise control over the act. While control over the act can be lost to the user if outputs are provoked by the user, this is not the case with simple prompts.
Case Citation
GEMA v OpenAI entities operating ChatGPT, Regional Court of Munich I, Judgment dated 11 November 2025, Case No 42 O 14139 24, GRUR 2025 1917.
Indian Kanoon link Not available as this is a German court decision. Visit date 11 March 2026.
Disclaimer
This case blog is based on the author’s understanding of the judgment. Understandings and opinions of others may differ. An AI application was used to generate parts of this case blog. Views are personal.