On 11/11, 2025, the Munich Regional Court handed a judicial victory to the music community with a milestone ruling that sent shockwaves through the artificial intelligence industry.
In a case brought by GEMA, Germany’s largest music rights organization, the court ruled that OpenAI’s ChatGPT violated copyright law by training its language models on protected song lyrics without permission. Specifically, GEMA pointed to nine song lyrics to make their case.
This decision represents the first time a European court has held an AI company directly liable for unauthorized use of copyrighted material. This sets a precedent that could reshape how AI companies operate across the continent.
It is not known yet whether OpenAI will appeal, so at this date the ruling is not yet legally binding.
Getting down to the brass tacks: Details of the GEMA lawsuit.
GEMA filed its lawsuit against OpenAI in November 2024, alleging that ChatGPT had memorized and could reproduce song lyrics from nine well-known German artists, including Herbert Grönemeyer, Helene Fischer, and Inga Humpe. The organization demonstrated that simple prompts could make ChatGPT spit out these lyrics almost verbatim—raising fundamental questions about copyright in the context of generative AI.
OpenAI argued that its language models don’t store specific training data but instead learn patterns and statistical correlations from datasets. The company contended that any reproduced text resulted from user prompts are beyond its control, essentially placing responsibility on individual users. OpenAI also invoked the EU’s text and data mining exceptions, claiming its training processes fell within legal boundaries.
The court, however, ruled that both the memorization of lyrics in the language models and their reproduction in chatbot outputs constitute copyright infringement under German law and the EU’s InfoSoc Directive. The judges found that OpenAI selected the training data, built and operated the system, and determined its architecture, making the company, not its users, responsible for the outputs.
A Breakthrough for creatives
What makes this ruling particularly significant is the court’s technical rigour. Drawing on scientific literature from information technology research, the judges accepted that training data can become embedded in model weights and remain retrievable, known as “memorization.” The court drew an analogy to MP3 compression, noting that it’s sufficient for a model to generate statistically probable sequences that recognizably reproduce copyrighted material. This is relevant even if individual data points can’t be isolated within the model.
The ruling rejected OpenAI’s claim that training AI models fell under text and data mining exceptions. While the court acknowledged that compiling training data might be covered by these provisions, it determined that the reproduction and storage of complete copyrighted works went beyond what the exception allows. The court emphasized that training AI is not an ordinary use to which authors have implicitly consented.
The court even found copyright infringement in a passage as short as 15 words, where creative elements concerning rhythm and grammatical structure were identifiable. This sets a low quantitative threshold for what constitutes actionable copying.
To everything (AI), turn, turn, turn.
As the song* goes, «There is a time to build up, a time to break down.» The implications of this ruling extend far beyond nine German songs. AI companies operating in the European Union now face significantly heightened legal risks if they cannot demonstrate that their training data was lawfully sourced.
Basically, what is built on copyrighted material can easily «break down» under judicial scrutiny. The court’s interpretation that AI models can constitute “copies” even when containing only probabilistic representations rather than exact text extends copyright law into uncharted territory.
For the AI industry, this decision poses both practical and philosophical challenges. The court pointed out that compliance may be technically difficult, as training data cannot easily be removed once embedded in a model. Nevertheless, it firmly placed the burden on providers to prevent further infringements. OpenAI has announced its intention to appeal, and the case will likely reach higher courts, possibly including the Court of Justice of the European Union.
Rights holders, meanwhile, are celebrating what GEMA CEO Tobias Holzmüller called “the first landmark AI ruling in Europe.” The decision strengthens arguments for collective licensing mechanisms similar to those long established in the music industry. GEMA has already developed an AI licensing model designed to let technology companies train on its catalogue legally while ensuring fair compensation for artists.
The financial stakes are brutal. While the damages awarded to GEMA remain undisclosed, they could reach hundreds of thousands of euros. Scaled across global operations and multiple jurisdictions, such liability could run into millions for AI companies that fail to secure proper licenses.
Tomorrow Never Knows
The model provided by GEMA, which also conceptually mirrors other agreements between rights-holders and AI developers should surely be established practice. The arrogant disregard of AI-proselytizers like OpenAI’s Sam Altman etc. for, well basically the whole history of human creation cannot become legal precedent.
This ruling aligns with broader European regulatory trends, particularly the EU AI Act’s emphasis on data governance and transparency. AI companies will likely need to maintain detailed records of datasets and demonstrate that appropriate licenses have been obtained—a significant operational burden.
The case also signals that Europe is charting its own course on AI regulation, distinct from the United States’ fair use doctrine, which could just as well be named «free use doctrine» under the current administration. With similar copyright lawsuits pending against OpenAI in Canada, Brazil, India, and multiple U.S. jurisdictions, the GEMA victory could inspire rights holders worldwide.
For now, the days of treating the internet as “a self-service buffet,” as GEMA’s Holzmüller put it, are ending for AI companies in Europe. Innovation must operate within the bounds of established copyright law, and that means obtaining licenses, paying royalties, and respecting the rights of human creators whose work powers these transformative technologies.
We’re only on the first verse of the epic «ballad of the struggle between the AI ogres and the copyrights holder knights», but this ruling represents an «intro» that might set the tone, rhythm and bpm for how it all unfolds in the coming few years!
*Turn Turn Turn by Pete Seeger.
