Incident Overview
On November 11, 2025, the Munich Regional Court issued Europe’s first substantive ruling on the legality of large language model (LLM) training methods, specifically targeting OpenAI’s ChatGPT. The court determined that the model’s ability to reproduce lyrics from German artists—including Herbert Grönemeyer—constituted unlawful copyright infringement. This ruling directly addressed the memorization of copyrighted material within the model’s weights and its subsequent reproduction in generated outputs, marking a critical precedent in AI and intellectual property law.
The case centered on whether OpenAI’s use of copyrighted song lyrics for training purposes violated German copyright law, particularly under the text-and-data-mining (TDM) exception. The court rejected OpenAI’s reliance on TDM defenses, establishing that the extraction and retention of creative works—such as lyrics—in the model’s training data and architecture amounted to infringement. This decision underscores the legal risks associated with unlicensed scraping and embedding of copyrighted content in AI systems.
While OpenAI has appealed the decision, the case sets a precedent that could influence future litigation and regulatory approaches to AI training practices across the European Union.
GEMA, the German society representing over 100,000 composers, lyricists, and publishers, played a pivotal role in the case. The organization’s database contains over 35 million songs and music works, making it a key stakeholder in disputes over AI and copyright.
Possible Root Cause
The root cause of the infringement, as determined by the Munich Regional Court, stems from OpenAI’s training methodology for ChatGPT, which involved the ingestion of vast quantities of copyrighted text data—including lyrics—without proper licensing or authorization. The court ruled that the model’s ability to reproduce specific lyrics upon user prompts demonstrated that the copyrighted material had been permanently encoded into the model’s weights during training.
Key factors contributing to the infringement include:
- This practice violated German copyright law, which requires explicit authorization for the reproduction or adaptation of protected works, even in automated systems.
- Memorization in Model Weights: The court found that the model’s architecture retained verbatim fragments of copyrighted lyrics within its parameters. This memorization enabled the model to generate near-identical reproductions of the original works, effectively replicating the infringing content in outputs. The court distinguished this from mere statistical pattern recognition, emphasizing that the model had internalized and stored protected material.
- Rejection of Text-and-Data-Mining Defenses: OpenAI argued that its use of the lyrics fell under the TDM exception, which permits the extraction of data from copyrighted works for non-commercial research purposes. However, the court dismissed this claim on two grounds: first, that TDM exceptions do not apply to commercial AI models like ChatGPT; second, that the reproduction of specific, identifiable lyrics in outputs constituted a direct infringement of the reproduction right under Article 15 of the German Copyright Act.
- Lack of Transformative Use: The court also rejected the notion that the model’s output constituted a transformative use of the original material. Unlike creative adaptations or parodies, ChatGPT’s reproduction of lyrics was deemed a verbatim replication, which does not qualify for fair use or fair dealing exceptions under German law.
The ruling clarifies that the mere presence of copyrighted material in training data—even if processed or transformed—can lead to liability if the model can regenerate protected content in a recognizable form.
Takeaways
The Munich court’s decision carries significant implications for AI developers, copyright holders, and regulators, shaping the future of LLM training and intellectual property law.
- Licensing is Non-Negotiable for Commercial AI Models: The court’s rejection of TDM defenses for commercial purposes establishes that AI developers must secure explicit licenses for copyrighted material used in training, even if the material is transformed or embedded in complex models.
- Memorization in Model Weights is Actionable: The ruling explicitly treats the storage of copyrighted material within a model’s weights as a form of reproduction under copyright law. This interpretation holds AI developers accountable not only for the data they ingest but also for how that data is structurally encoded in the model.
- Reproducibility Triggers Liability: The ability of an AI model to generate recognizable fragments of copyrighted works—such as song lyrics—creates legal exposure. The court’s focus on ChatGPT’s output capabilities suggests that even indirect or partial reproductions may be actionable. This shifts the burden onto developers to ensure that their models cannot regenerate protected content, even if the training process itself is opaque.
- Regulatory Precedent for the EU: As the first substantive EU ruling on AI and copyright, the Munich decision sets a benchmark for future cases. Courts in other EU jurisdictions may cite this ruling to interpret similar disputes, particularly those involving creative industries. The decision also signals that national courts may adopt stricter stances on AI training than those taken by regulators or industry self-governance bodies, such as the EU AI Act’s current provisions on copyright compliance.
- Appeal Process and Uncertainty: While the ruling is legally binding, OpenAI’s appeal introduces a period of uncertainty. The outcome of the appeal could either reinforce the court’s interpretation or prompt a reassessment of how TDM exceptions apply to AI. Until the appeal is resolved, developers operating in Germany—or exporting models to German users—face heightened legal risks. The case also raises questions about whether other jurisdictions will follow suit, particularly those with strong copyright protections like the UK or France.
- Broader Impact on AI Ethics and Transparency: The ruling indirectly addresses concerns about the opacity of AI training processes. By holding OpenAI liable for the model’s outputs, the court implicitly recognizes that developers must exercise greater control over what their models learn and reproduce.
- Economic Consequences for AI Developers: The financial implications of the ruling extend beyond legal penalties. Additionally, the case could deter investment in unlicensed training datasets, particularly in creative sectors where copyright enforcement is robust.
- Shift in Industry Practices: The decision is likely to prompt AI developers to adopt more cautious approaches to data sourcing.
- Strengthened Position for Rights Holders: For creators, publishers, and CROs, the ruling validates their efforts to protect intellectual property in the digital age. It sends a clear message that AI development must respect existing copyright frameworks, rather than operating under the assumption that scale or transformation justifies unlicensed use. This could embolden rights holders to pursue similar cases against other AI developers, particularly those training on large corpora of books, music, or visual art.
The Munich court’s decision marks a turning point in the intersection of AI and copyright law, with ripple effects across technology, entertainment, and legal sectors.