On November 11, 2025, the Munich I Regional Court (Landgericht München I) issued a landmark ruling in GEMA v. OpenAI (Case No. 42 O 14139/24), holding OpenAI liable for copyright infringement involving the use of song lyrics to train and operate ChatGPT. The case began when Gesellschaft für musikalische Aufführungs- und mechanische Vervielfältigungsrechte (GEMA), Germany’s principal music rights society, alleged that OpenAI had used lyrics from nine well-known German songs in its training data and that ChatGPT could reproduce substantial portions of those lyrics upon request. After technical analysis and expert testimony, the court agreed with GEMA, concluding that the model had effectively “memorized” the lyrics and was capable of outputting them in a way that constituted a copyright-relevant reproduction.
In its reasoning, the court held that the internal encoding of copyrighted lyrics within a model’s parameters amounts to reproduction under Section 16 of the German Copyright Act (UrhG) and Article 2 of the EU InfoSoc Directive, even though the works are not stored as conventional files. The ability of ChatGPT to output recognizable and substantial parts of those lyrics constituted an additional act of reproduction and, because the outputs are provided to users over the internet, an act of making the works available to the public under Section 19a UrhG and Article 3 of the InfoSoc Directive. The judgment therefore treats both the internal model training and its public-facing responses as infringing uses, marking one of the clearest judicial statements to date on how copyright law applies to large language models.
One of the most significant aspects of the ruling is its treatment of the EU’s text-and-data-mining (TDM) exception. The court rejected OpenAI’s argument that the ingestion of copyrighted lyrics was permissible under Section 44b UrhG, Germany’s implementation of the EU’s TDM provisions. Although the court acknowledged that preparatory copies for analytical purposes may fall under the exception, it held that the memorization of protected works in the model, as well as the possibility of reproducing them for the public, interferes with the rights-holder’s legitimate economic interests and therefore lies outside the scope of the TDM exception. This interpretation significantly narrows the protective effect of TDM for AI developers (at least in Germany) in situations where a model can output recognizable copyrighted text.
Equally important is the court’s rejection of the argument that the responsibility for any infringing output lies with the user who prompts the model. The court found that OpenAI, not its users, selects and curates the training data, designs the model architecture, and controls the technical mechanisms that enable the model to generate responses. As a result, OpenAI bears primary responsibility for ensuring that its models do not memorize or emit copyrighted works without authorization. This finding places the burden of avoiding copyright infringement under German law squarely on AI providers and is likely to influence how indemnities, warranties, and dataset disclosures are negotiated in AI procurement contracts across the EU.
As relief, the Munich court granted GEMA a cease-and-desist order, required OpenAI to provide information and disclosures regarding the use of the lyrics, and acknowledged a right to damages or license fees. Although the exact monetary award has not yet been made public, reporting suggests that it may reach significant levels. OpenAI has publicly stated that it disagrees with the decision and is considering an appeal.
This ruling is poised to have ripple effects across the European Union. It is already being described as the EU’s first major judicial decision on AI and music copyright, with rights-holders and regulators closely watching its implications. The decision signals that European courts may take an expansive view of what constitutes reproduction in the context of AI, recognizing not only direct outputs but also the internal technical processes through which machine-learning models encode and store protected works. It also suggests that rights-holders may increasingly pursue claims not only for public-facing outputs but also for training-stage uses that embed copyrighted works within model parameters.
Taken together with emerging transparency, data-governance, and model-documentation requirements under the EU Artificial Intelligence Act, the GEMA v. OpenAI decision underscores a broader regulatory shift: AI developers in the EU will be expected to obtain appropriate licenses for copyrighted training data or implement strong technical safeguards to prevent memorization and extraction of protected content. European collecting societies have already signaled that this ruling may pave the way for more formalized licensing frameworks for AI training. As AI systems continue to rely on large quantities of copyrighted material, the Munich judgment may become a foundational precedent in the EU, shaping how generative AI models are trained and deployed for years to come.