A landmark decision
On 11 November 2025, the Regional Court of Munich I ruled that OpenAI’s language models unlawfully used song lyrics by nine German authors. GEMA, acting as the collecting society, claimed that OpenAI’s models had memorised and reproduced copyrighted lyrics without permission. Comparable collecting societies across Scandinavia – such as KODA (Denmark) play similar roles in protecting authors’ rights and are closely watching this development.
Memorisation equals reproduction
The Court found that both memorisation within GPT-4 and GPT-4o and the reproduction of lyrics in chatbot outputs constitute unauthorised duplication under German and EU copyright law. Even if stored as probability values, the models’ parameters were considered a “reproduction in any form.”
This interpretation is significant because it treats AI model parameters as sufficiently concrete to qualify as copies – an approach that could radiate across the EU.
No text and data mining defence
OpenAI’s reliance on the text and data mining (TDM) exception was rejected. The judges held that the exception covers temporary analysis, not lasting reproductions embedded in a model. Extending it by analogy was deemed impermissible, as memorisation interferes with exploitation rights.
The Danish Ministry of Culture’s Expert Group on Copyright and Generative AI recently echoed similar concerns, noting that the current EU TDM exemptions were never designed for large-scale, irreversible memorisation within neural networks. The group highlighted that lasting parametric storage may fall outside the scope of lawful TDM, reinforcing the Munich court’s position.
Operator responsibility
According to the Court, OpenAI – not its users – is responsible for infringing outputs, since it designs and trains the models that generate them. Claims based on moral rights were dismissed.
Implications
The ruling is the first in Europe to treat AI memorisation as a copyright-relevant act. If upheld, it could require AI developers to seek licenses for training data and reshape how copyright applies to generative AI.
For Danish rights holders and societies such as KODA, the judgment signals that similar arguments could gain traction in their jurisdictions, especially as several Scandinavian legal bodies have recently emphasised the need to clarify whether AI training should require licenses. The Danish Ministry of Culture’s Expert Group on Copyright and Generative AI, in particular, has recommended exploring a clearer licensing framework to ensure legal certainty for both creators and AI developers.
At aera, we see this decision as a key moment for copyright and AI. It highlights the need for compliance strategies, licensing frameworks, and technical safeguards to reduce risk.
Do you need help?
aera advises clients on the full spectrum of IP and AI-related issues, from copyright compliance and data governance to licensing and regulatory strategy. We are glad to assist in evaluating the implications of this decision for your AI projects or data-driven business models.
