Today, the Munich Regional Court (Landgericht München I) delivered its ruling in the significant GEMA v. OpenAI case (case number 42 O 14139/24). We are awaiting the detailed written reasoning, but the main points from the court’s press release are now clear and they have far-reaching implications.
The court’s key conclusions are as follows:
First, the court ruled that when specific song lyrics are reproduced almost identically by a large language model in response to simple prompts, this constitutes evidence that the model has memorised those lyrics in its parameters. The court also ruled that this memorisation constitutes reproduction of the copyrighted work within the model itself (thus the reproduction claim is not limited to outputs). Furthermore, the court determined that this memorisation falls outside the text and data mining exceptions provided by Directive (EU) 2019/790. Collectively, these three conclusions establish that LLM training practices incorporating copyrighted material, even when seemingly carried out in accordance with TDM exceptions, do not automatically shield providers from liability for copyright infringement when the resulting model memorises and reproduces protected works.
These conclusions — that memorisation is reproduction, that TDM exceptions do not shield memorisation and that reproduction without a licence is infringement — create a legal framework that, if adopted and applied by courts in other EU member states, could fundamentally change how AI developers and providers should approach copyrighted material and how authors can protect their works.
This would mean:
For authors and rights holders:
Memorisation in LLM parameters constitutes a form of reproduction that triggers full copyright protection across the European Union. The reasoning behind the Munich court ruling establishes that authors’ works incorporated into model parameters should not be treated as mere mathematical abstractions or pattern extraction, but rather as actual reproductions of their copyrighted material. Consequently, authors retain their right to control and monetise the use of their works.
The Munich court’s decision effectively rejects the argument that no reproduction has occurred simply because LLMs do not store works in a retrievable database format.
Additionally, the ruling implies that authors or collecting societies such as GEMA can now enforce protection and/or licensing requirements across multiple EU jurisdictions, since the damage (manifested through infringing outputs) occurs in every Member State in which the model and its outputs are accessible. This creates a cascading enforcement mechanism whereby any EU Member State court to which the outputs are made public could theoretically exercise jurisdiction to award damages and issue injunctions limited to that territory.
For LLM developers and training practices (including fine-tuning)
Developers can no longer argue that training on copyrighted material constitutes permitted text and data mining when the resulting model memorises and reproduces identifiable portions of that material.
Although the EU TDM exceptions in Directive (EU) 2019/790 allow the commercial use of copyrighted works for data mining purposes when rights holders have not opted out, the Munich court’s reasoning introduces an important caveat: the TDM exception does not protect activity resulting in the verbatim or near-verbatim reproduction of protected works. This distinction is crucial. It means that AI developers and providers must implement robust technical measures to prevent memorisation during training, rather than merely mitigating output infringement.
The burden now substantially shifts towards developers and providers to demonstrate that their training procedures either operated on opted-out material with proper verification of such opt-outs or employed architectures and training protocols that mathematically prevent memorisation of the source material. This will probably necessitate architectural innovations, such as differential privacy techniques or federated learning approaches that restrict access to training data and limit the capacity of models to memorise specific sequences.
The ruling creates particular complications for the practice of fine-tuning pre-trained models. When developers fine-tune an existing model using copyrighted material, they inherit the potential liability of the base model if it contains memorised works. Additionally, if the fine-tuning process introduces new memorisation of copyrighted material, separate infringement liability attaches to that activity.
Consequently, companies offering fine-tuning services or engaging in custom model adaptation must now conduct due diligence on both their fine-tuning datasets and the provenance and potential memorisation patterns of their base models. If the base model has been trained on unlicensed copyrighted material in a way that results in memorisation, specialised fine-tuning cannot remedy that initial infringement.
Implications for application of the TDM exception in Directive 2019/790
The Munich decision implicitly interprets the scope of the TDM exceptions in Directive 2019/790, although it applied German national rules resulting from the transposition of this Directive. While these exceptions theoretically permit the commercial reproduction and extraction of copyrighted works for data mining purposes, subject to an opt-out mechanism that allows rights holders to reserve their rights, the court’s reasoning suggests that TDM exceptions do not immunise activity that produces verbatim output of training data.
In other words, TDM permits the analytical process of examining copyrighted materials to extract patterns and correlations, but does not permit memorisation and reproduction of those materials in their original form.
Therefore, even if a rights holder failed to opt out, a developer’s failure to implement memorisation prevention measures could still constitute an infringement of the underlying copyright.
Jurisdictional implications under the Brussels I bis Regulation and
A single AI model could face simultaneous proceedings in different EU Member States, each of which could issue territorially limited decisions based on their national copyright laws.
Article 7(2) of the Regulation allows copyright claimants to sue “where the harmful event occurred or may occur”. This refers to both (1) the location where the infringing act originates and (2) every location where the damage manifests.
Copyright infringement disputes can thus generally be litigated anywhere the infringement causes damage, meaning rights holders in the EU can usually initiate proceedings in their home jurisdiction.
The Munich court’s ruling is also interesting when it comes to the application of the lex loci protectionis principle (Article 8 of the Rome II Regulation). It does not define the infringement and the resulting applicable substantive copyright laws by the activity of training the model (such an approach would likely lead to the application of US copyright law and not German copyright law). Instead, it likely relates to the resulting status and qualities of the model and the fact that the model has been made available in Germany. However, the details of its approach will be known after the publication of the integral written judgment.
