In brief
- Judges found GEMA’s claims valid, ordering OpenAI to cease reproduction and provide damages and disclosure.
- The court said GPT-4 and GPT-4o “memorized” lyrics, amounting to reproduction under EU copyright rules.
- The decision, not yet final, could set a major European precedent on AI training data.
Germany’s national music rights organization secured a partial but decisive win against OpenAI after a Munich court ruled that ChatGPT’s underlying models unlawfully reproduced copyrighted German song lyrics.
The ruling orders OpenAI to cease reproduction, disclose relevant training details, and compensate rights holders. It is not yet final, and OpenAI may appeal.
If upheld, the decision could reshape how AI companies source and license creative material in Europe, as regulators weigh broader obligations for model transparency and training-data provenance.
The case marks the first time a European court has found that a large language model violated copyright by memorizing protected works.
In its decision, the 42nd Civil Chamber of the Munich I Regional Court said that GPT-4 and GPT-4o contained “reproducible” lyrics from nine well-known songs, including Kristina Bach’s “Atemlos” and Rolf Zuckowski’s “Wie schön, dass du geboren bist.”
The court held that such memorization constitutes a “fixation” of the original works in the model’s parameters, satisfying the legal definition of reproduction under Article 2 of the EU InfoSoc Directive and Germany’s Copyright Act.
“At least in individual cases, when prompted accordingly, the model produces an output whose content is at least partially identical to content from the earlier training dataset,” a translated copy of the written judgement provided by the Munich court to Decrypt reads.
The model “generates a sequence of tokens that appears statistically plausible because, for example, it was contained in the training process in a particularly stable or frequently recurring form,” the court wrote, adding that because this “token sequence appeared on a large number of publicly accessible websites“ it meant that it was “included in the training dataset more than once.”
In the pleadings, GEMA argued that the model’s output lyrics were almost verbatim when prompted, proving that OpenAI’s systems had retained and reproduced the works.
OpenAI countered that its models do not store training data directly and that any output results from user prompts, not from deliberate copying.
The company also invoked text-and-data-mining exceptions, which allow temporary reproductions for analytical use.
“We disagree with the ruling and are considering next steps,” a spokesperson for OpenAI told Decrypt. “The decision is for a limited set of lyrics and does not impact the millions of people, businesses, and developers in Germany that use our technology every day.”
OpenAI claims systems like theirs do not store or contain training data and thus do not hold copies of lyrics or other texts. Instead, these models learn patterns and generate new outputs based on patterns, OpenAI said.
The company told Decrypt that treating a model as if it contains stored works reflects a misunderstanding of how the technology works.
The court rejected those defenses, ruling that full reproductions embedded in a model’s structure fall outside the scope of data-mining exemptions.
“Training the models is not to be regarded as a usual and expected form of use that the rights holder must anticipate,” the court wrote. “This applies all the more when—as in the present case—the works are reproduced in the model, something that even the defendants themselves consider undesirable and against which countermeasures are taken.”
Decrypt reached out separately to GEMA for comment but has yet to receive a response by press time.
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
