Munich AI ruling in summary

Landgericht München I, the Munich regional court, has handed down a judgment in the case GEMA v OpenAI which will heavily impact the debate revolving around AI and music. Even if this judgment does not apply directly in Switzerland, it is a good indication of how courts in Europe might approach usage by AI systems of musical works. The judgment is not yet final and will probably be referred to higher instances.

What it was about
GEMA established that, with very simple prompts, ChatGPT could output the complete lyrics (e.g. “Männer” by Herbert Grönemeyer) of well-known songs. Parts of the lyrics were reproduced word for word, without any licence. In GEMA’s view, this showed that the works in question had been used in training the AI models and that they had not just been analysed, but permanently memorised in the system. Consequently, GEMA demanded an injunction and damages, and instituted proceedings with the regional court in Munich.

OpenAI confirmed that unintentional repetitions of training materials could occur, but explained that the models are basically designed only to learn patterns and statistical correlations. At the same time, the company claimed that the output depended on user input and was not directly based on stored training data.

How the Court decided
OpenAI may not retain protected song lyrics in the model or reproduce them in ChatGPT answers. Moreover, the company must indemnify GEMA for the damage incurred and disclose which data was used for training and what revenues were generated with the model. GEMA had also claimed an infringement of personal rights. However, the Court rejected this claim.

Why memorisation in the model is decisive
A crucial point of the judgment was the determination that AI models can actually retain parts of the training data. Lyrics do not simply disappear in the learning process but can be memorised in the system and show up later. For the Court, such memorisation qualifies as use of protected content. Whether the underlying technology is fully understood or not is irrelevant. The decisive factor is that the content is in the model.

Why the TDM rule does not apply
The EU has an exception for text and data mining (TDM). This is designed to enable large volumes of data to be analysed. The Court stated clearly that this rule only applies to the assembly of analysis material and does not extend to permanent storage of parts of works in the trained model. The memorisation of lyrics no longer serves analytical purposes and is not, therefore, covered by this exception.

What applies for consent
The Court underscored that there was no reason why authors should automatically expect their works to be used for training larger AI models since that was neither a customary nor a predictable usage. Neither is the notion that AI training has meanwhile become standard practice in the industry an acceptable justification. Furthermore, the Court pointed out that the operators of such systems themselves actually try to avoid undesired storage.

Who is responsible
The judgment clearly states that OpenAI is responsible for the release of the model since operators are responsible for the architecture of their models and the memorisation of training data. A person who simply enters a prompt triggers an answer but does not determine the content of such answer. The content comes from the model itself, and is therefore the responsibility of the provider.

Conclusion
This case shows that the memorisation of work content in AI models remains the key issue. The retention and output of training material by models is increasingly considered a relevant use. If higher instances, possibly also the European Court of Justice, confirm that such storage within models qualifies as use, this could mitigate the problematic fact that most models are trained abroad.
As a result, transparency – in the matter of training data especially – will become even more important for Switzerland because it is the only way to determine whether protected music is involved and a license required. Even without direct effect, this judgment is indicative of how courts could adjudicate such cases in the future.