In November 2024, the German collecting society GEMA initiated proceedings against OpenAI, marking Europe’s first major copyright lawsuit involving generative AI.
In this lawsuit, GEMA claims that OpenAI has reproduced the lyrics of the repertoire of its 100.000 members to train its Large Language Model (LLM) ChatGPT, without obtaining a license or paying any remuneration. This despite being informed of this obligation by GEMA. According to GEMA, the model reproduces original song lyrics in response to simple user prompts, proving that the system memorized these works during training.
OpenAI rejected GEMA’s allegations, arguing that its LLM does not store or copy training data but merely generates responses based on user prompts. The company asserted that users, not OpenAI, are responsible for any outputs and that any potential infringements fall under the text and data mining exception in EU copyright law.
According to GEMA’s review of the oral hearings, the case attracted ‘huge interest’, with prominent German artists.
The Munich Regional Court confirmed that lyrics from nine well-known German artists were reproduced. According to the Court, the LLM memorized these lyrics. The Court compared the training data with the model’s outputs. Based on this comparison, the Court ruled out that coincidence as cause of reproduction, given the complexity and length of the lyrics. The court concluded that the model did not just extract information but incorporated a complete copy of the lyrics into its parameters.
According to the Court, the memorization entails that the lyrics are stored in the model in a way that makes them reproducible. This constitutes an embodiment, which is the prerequisite for reproduction under copyright law. The Court states that, according to CJEU case law, it is sufficient that the copyrighted work can be perceived using technical means to conclude that the song lyrics are reproduced within the model.
The Court finds that such memorization does not qualify as text and datamining. Even if the lawmakers did not anticipate on memorization when drafting the law, this should not fall under this exception. The exception namely only aims at covering preparatory acts that do not interfere with exploitation of the works and violate the rights of rightsholder. Other exceptions were also deemed inapplicable by the Court.
Lastly, the Court held OpenAI, and not the users, responsible for the reproductions. OpenAI selects the training data, trains and operates the models. They are thus responsible for the models’ architecture and memorization of training data.
As of now, the judgment is not final yet and remains open to appeal.
