On 11 November 2025, the Landgericht München I (case 42 O 14139/24) held that OpenAI's GPT-4 and GPT-4o models infringe German copyright law by memorising and reproducing song lyrics administered by GEMA, finding violations of paragraphs 15, 16 and 19a of the Urheberrechtsgesetz (UrhG) and rejecting OpenAI's reliance on the text and data mining (TDM) exception under paragraph 44b UrhG. The case is now on appeal, but it already signals that German courts are prepared to scrutinise AI training pipelines with considerable rigour – and that the practical and financial stakes for any organisation deploying or licensing large language models in Germany are material.
The Legal Framework: German Copyright Law and AI Training
German copyright protection for musical works – including lyrics – arises automatically upon creation and vests in the author under the UrhG. The exclusive rights engaged in the GEMA proceedings are the reproduction right (§ 16 UrhG), the making-available right (§ 19a UrhG) and the general exploitation right framework of § 15 UrhG. These provisions implement the EU InfoSoc Directive and are broadly aligned with the copyright frameworks of other EU Member States, though German courts have historically interpreted the reproduction right with particular breadth.
The TDM exception introduced by § 44b UrhG – transposing Article 4 of the EU Digital Single Market (DSM) Directive – permits text and data mining for any purpose, subject to one crucial carve-out: rights holders may opt out. GEMA had done precisely that. The München I court found that OpenAI had trained on material covered by a valid opt-out, making the exception unavailable. This reasoning underscores a structural tension at the heart of EU TDM law: the opt-out mechanism shifts the compliance burden onto AI developers to monitor and respect rights-holder reservations, a task that becomes progressively more complex as training datasets grow in scale and provenance becomes harder to audit.
The research TDM exception under § 60d UrhG, which does not permit opt-out, was not considered applicable to OpenAI's commercial activities. German law therefore currently offers no general safe harbour for commercially motivated AI training on copyright-protected content where a valid opt-out exists.
The München I Ruling: What the Court Actually Decided
The court's central finding on 11 November 2025 was that GPT-4 and GPT-4o had memorised specific song lyrics to a degree constituting reproduction within the meaning of § 16 UrhG. The court rejected the argument that model weights represent a fundamentally different kind of artefact from a stored copy of a text. In doing so, it declined to follow the analysis adopted by the English High Court in Getty Images v Stability AI [2025] EWHC 2863 (Ch) (4 November 2025), where the Chancery Division found that model weights do not store copies and that secondary infringement claims had failed – a significant doctrinal divergence across two major European jurisdictions that practitioners advising multinational clients must account for carefully.
Three points from the München I judgment warrant particular attention:
- Memorisation as reproduction: The court linked the act of memorisation during training directly to the reproduction right, not merely to the output. This potentially exposes AI developers to liability at the training stage, before any end-user query is made.
- Opt-out effectiveness: The court treated GEMA's published opt-out as legally operative and enforceable. Developers who cannot demonstrate that they checked and respected opt-outs at the time of data ingestion face a structural gap in their defence.
- Scope of § 19a UrhG: The making-available right was engaged by the model's capacity to reproduce lyrics on demand – suggesting that even probabilistic output capability, rather than actual dissemination, may suffice to trigger infringement.
OpenAI is appealing the decision. The appellate proceedings before the Oberlandesgericht München will likely address whether the lower court's approach to model weights is consistent with the DSM Directive's recitals and with CJEU jurisprudence. Until a final appellate ruling is issued, the first-instance judgment remains the most authoritative statement of German law on this point.
If your organisation is training, fine-tuning or deploying AI models that ingest European content, we would welcome the opportunity to discuss the implications for your specific operations. Please contact us at info@ferrazwhitmore.com or visit our contacts page.
The Territorial Question: Does Training Outside the EU Matter?
A second München I ruling – reported on 31 July 2026 and not yet final – indicated that training on protected musical works may infringe German copyright even where the training process took place outside the EU. This is a reported, non-final indication rather than settled law, and it should be treated with corresponding caution. Nevertheless, it signals that at least one German court is willing to consider a form of effects-based jurisdiction, analogous to the approach taken in competition law and certain data protection contexts, where the place of harm or the place of making-available to German users anchors jurisdiction regardless of where the technical act of training occurred.
If this reasoning is affirmed on appeal, it would have far-reaching consequences for AI developers headquartered outside Europe who rely on the geographic separation of training infrastructure as a shield against EU liability. It would also create significant tension with US law, where training practices that are contested in Germany may enjoy different treatment under the fair use doctrine. Cross-border IP litigation strategies would need to account for this divergence explicitly.
For context, the English High Court in Getty Images v Stability AI [2025] EWHC 2863 (Ch) took a notably different path on the storage and reproduction questions, finding that model weights do not constitute stored copies of training images – a conclusion that, if applied to lyrics, would cut off liability at the training stage. The doctrinal gap between the German and English approaches is not merely academic: it creates genuine forum-shopping incentives and complicates any pan-European licensing or litigation strategy. Our analysis of the UK dimension is available at Getty v Stability AI – United Kingdom.
The CJEU Horizon: Like Company v Google (C-250/25)
The pending CJEU reference in Like Company v Google (C-250/25), heard on 10 March 2026 with an Advocate General opinion expected in September 2026, may provide important clarification on the scope of the DSM Directive's TDM provisions and on the conditions under which an opt-out must be communicated and respected. The case arises in a different factual context but touches on several of the interpretive questions left open by the München I reasoning – in particular, what technical and procedural steps a data user must take to identify and honour opt-outs at scale.
Depending on the AG opinion and any subsequent CJEU ruling, German courts hearing future AI copyright cases may need to revise their approach to § 44b UrhG. Practitioners advising clients currently in proceedings or in licensing negotiations should monitor C-250/25 closely, as it could shift the legal baseline before appellate proceedings in the GEMA case are concluded.
Practical Risks for International Clients Operating in Germany
The GEMA v OpenAI proceedings crystallise several categories of risk that any organisation developing, licensing or integrating AI systems in the German market should assess:
- Training data provenance: Organisations that cannot demonstrate a documented audit trail of their training datasets – including evidence of opt-out checking at the time of ingestion – face an elevated litigation risk under the current German framework. This applies equally to companies using third-party models if they have contractually assumed responsibility for compliance.
- Copyright infringement exposure in Germany: IP litigation before the German courts can be materially costly and, in cases involving collecting societies such as GEMA, may involve claims aggregated across large repertoires. The Bundesgerichtshof has historically been willing to uphold significant damages awards in copyright matters.
- Interaction with the AI Act: The EU AI Act requires providers of general-purpose AI models to publish and maintain a sufficiently detailed summary of the content used for training. Non-compliance with this obligation – now applicable to models with systemic risk designations – may interact with copyright infringement claims by facilitating rights-holder identification of infringing uses. Organisations should consider AI Act compliance and copyright risk as linked rather than separate workstreams; our dedicated analysis is available at AI Act Compliance – Germany.
- Design registration and trade mark dimensions: Although the GEMA proceedings concern musical copyright, broader AI training datasets may also implicate protected designs and registered trade marks. Where model outputs reproduce distinctive visual marks or registered designs, separate claims under German and EU design and trade mark law may arise alongside copyright infringement actions.
- Licensing negotiations: The existence of the München I judgment strengthens the negotiating position of rights-holder organisations. Developers who previously declined collective licensing arrangements may find that the cost-benefit calculus has shifted, particularly given the territorial reach questions raised by the second München I indication.
If you are assessing litigation exposure or structuring a licensing strategy in response to these developments, our IP and litigation and arbitration team in Germany is available to advise. You can reach us at info@ferrazwhitmore.com.
Strategic Recommendations
In light of the current German legal landscape, the following strategic considerations are relevant for organisations with AI exposure in Germany:
- Conduct a training data audit: Review the provenance of datasets used to train or fine-tune any model deployed in Germany. Identify whether opt-outs were published by any rights holders whose material was ingested, and document the steps taken – or not taken – to identify and honour those reservations.
- Review contractual allocation of IP risk: Where an organisation licenses a third-party model rather than training its own, examine whether the licence agreement allocates copyright infringement risk adequately. The München I reasoning could expose a downstream licensee to claims where the licensor's training practices are later found to have been infringing.
- Engage with collective licensing frameworks: In sectors with strong collective rights management – music, publishing, visual arts – proactive engagement with licensing bodies offers a route to legal certainty that litigation cannot provide. A negotiated licence defines scope, territory and remuneration without the unpredictability of ongoing proceedings.
- Monitor the appellate proceedings and C-250/25: The legal position is not settled. The Oberlandesgericht München may affirm, modify or reverse the first-instance reasoning, and the CJEU reference may in turn inform that court's analysis. Strategic decisions made before appellate clarity carry a higher degree of legal uncertainty.
- Integrate AI Act and copyright compliance: Given the transparency obligations under the AI Act and their practical interaction with rights-holder claims, a unified compliance framework is more efficient and legally coherent than addressing each regime in isolation.
Outlook
The GEMA v OpenAI case marks a significant moment in the German and broader European legal treatment of AI training data. The first-instance judgment demonstrates that German courts are willing to apply established copyright doctrine to AI systems without waiting for legislative clarification – a disposition consistent with the Bundesgerichtshof's tradition of purposive, rights-protective interpretation. The divergence from the English High Court's approach in Getty Images v Stability AI illustrates that there is no single European answer to these questions yet, and that the resolution may ultimately require harmonisation at the CJEU level.
For international clients, the practical message is that Germany cannot be treated as a jurisdiction where AI copyright risk is theoretical or remote. The combination of active collecting society enforcement, a court willing to find training-stage infringement, and a pending territorial extension question creates a risk profile that warrants serious legal attention now, rather than after proceedings have been issued.
We are happy to discuss how these developments affect your organisation's operations in Germany and across the EU. Please write to us at info@ferrazwhitmore.com or visit our contacts page to arrange a consultation.
Disclaimer: This publication is provided for informational purposes only and does not constitute legal advice. The information herein should not be relied upon as a substitute for professional legal counsel tailored to your specific circumstances. Ferraz & Whitmore assumes no liability for actions taken or not taken based on the contents of this material. For advice regarding your particular situation, please contact info@ferrazwhitmore.com.