On 11 November 2025, the Civil Chamber of the Munich Regional Court issued its landmark ruling in GEMA v OpenAI (Case No. 42 O 14139/24). The ruling represents the European Union’s first judicial determination on whether the use of copyrighted material for AI training constitutes unlawful reproduction. This case was brought by GEMA, a music rights society, and concerned nine well-known German songs, including “Atemlos” by Kristina Bach and “Wie schön, dass du geboren bist” by Rolf Zuckowski, whose lyrics were allegedly memorised in OpenAI’s GPT-4 and GPT-4o models and reproduced in ChatGPT outputs.[1]
The court ultimately held that OpenAI both reproduced the protected lyrics during training and communicated them to the public through ChatGPT outputs.[2] This result has implications not only for copyright, but for the broader political narrative around Europe’s regulatory stance toward AI development.[3] The ruling positions the EU as a jurisdiction willing to apply copyright law stringently to AI developers, in contrast with approaches in the US and UK.[4]
Before GEMA v OpenAI, German courts had already touched on AI-related copyright questions in Kneschke v LAION, a case centred on text and data mining (“TDM”).[5] As noted in a previous analysis,[6] the Hamburg court held that a rights-holder’s opt-out does not always need to be expressed in machine-readable form; in certain contexts, a clear natural-language reservation of rights may be sufficient.[7] Kneschke v LAION, however, concerned dataset creation rather than output-level reproduction. Its relevance here is mainly contextual: it shows that German courts have started to grapple with the copyright implications of AI training, even though GEMA v OpenAI raises distinct issues around memorisation and the reproduction of protected works.
Memorisation as Reproduction?
A central element of the court’s reasoning was the concept of “memorisation,” which it treated as a question of fact rather than abstraction.
The court accepted GEMA’s evidence that song lyrics were indisputably part of the training data, could be reproduced by simple prompts, and appeared recognisably in outputs, thereby demonstrating memorisation.[8] In the court’s view, it was irrelevant whether the lyrics were stored verbatim or represented statistically; the decisive consideration was that they were embodied in model parameters and extractable through ordinary user input.[9] This broad interpretation led the court to deem memorisation in this particular case a form of reproduction under Section 16 UrhG and Article 2 of the InfoSoc Directive, which define reproduction to include any fixation enabling perceptibility through technical means.[10] The judgment extended the meaning of “storage” in the digital realm to encompass probabilistic model weights, thereby equating statistical correlations with copies of protected works.[11] This approach challenges the technical argument advanced by OpenAI that its models do not “store” works but only derive correlations, a position the court rejected as insufficient to avoid liability.[12]
Limits of the Text and Data Mining Exceptions
Both sides debated whether AI training falls within the EU’s text and data mining framework.
The court acknowledged that machine learning is generally covered by TDM exceptions but concluded that memorisation of entire or substantial parts of works does not constitute mere analysis of information.[13] According to the court, TDM allows transient, analytical copies but does not permit permanent embodiment of works in trained AI models.[14] Moreover, because the disputed lyrics reappeared in outputs, the court held that such uses exceeded the scope of §44b UrhG and Articles 3–4 of the DSM Directive.[15] The court refused interpretative flexibility aimed at supporting innovation, reasoning that authors’ economic rights would be prejudiced by allowing memorisation within models to fall under the TDM exceptions.[16] This restrictive reading significantly narrows the safe harbour AI developers had hoped to rely upon and signals that unlicensed ingestion of protected works is presumptively unlawful in the EU.[17]
Output-Level Infringement and the Allocation of Responsibility
The judgment also held that output reproduction constituted independent copyright infringement. ChatGPT’s ability to reproduce lyrics verbatim or with slight variations amounted to acts of reproduction and making available to the public under Sections 16 and 19a UrhG.[18] The court rejected OpenAI’s claim that the end-user should bear responsibility, instead holding that liability attaches to the model provider because it selects the training data, designs the architecture, and chooses to deploy the system.[19] This aligns with the court’s broader reasoning that innovation-driven liability exemptions applicable to early internet platforms are no longer justified for large AI developers.[20] The German ruling departs from recent UK precedent, where in Getty Images v. Stability AI the High Court declined to find that model weights contained stored copies of protected works.[21] While the UK court emphasised territorial limits and absence of verbatim storage, the Munich court applied a broader theory of embodiment and treated memorisation as sufficient to establish liability.[22]
Consequences for AI Developers
The judgment imposes extensive obligations on OpenAI, including an injunction against further reproduction of the lyrics in both the model and its outputs.[23] The court recognised that removing training data from an existing model is technically difficult and may be impossible; nonetheless, responsibility remains with the provider to avoid future infringements.[24] This may require output filters, retraining, or licensing, signalling that blanket licensing of training data could become the normative expectation for EU-compliant AI development.[25]
The court also ordered OpenAI to cease storing unlicensed song lyrics on German infrastructure, a remedy extending beyond output controls and targeting the model itself.[26] Given OpenAI’s assertion that “there is no unlearning,” the company may face the extreme possibility of disabling certain model versions in Germany unless it negotiates licences.[27] The ruling interacts with Article 53 of the AI Act, which obliges GPAI developers to ensure training data complies with EU copyright law irrespective of where training occurred, thereby reinforcing licensing as a structural requirement.[28]
Broader Significance
Although appeals are expected, the decision contributes to a growing body of German case law confronting the intersection of generative AI and copyright. While GEMA v OpenAI is only the second major judgment after Kneschke v LAION, Germany appears to be playing an outsized early role in articulating the legal boundaries of AI training.
The ruling also highlights broader political dynamics: the EU’s tendency to prioritise robust rights-holder protection, in contrast with the more permissive or innovation-oriented approaches seen in the US.[29] As further cases – including the Like Company referral concerning Google’s Gemini chatbot[30] – reach the CJEU,[31] it remains to be seen whether the Munich court’s expansive theory of memorisation will become the European standard.
[1] Hembt, Dr. Simon. “Landmark Ruling of the Munich Regional Court (GEMA v OpenAI) on Copyright and AI Training.” Bird & Bird, 14 Nov. 2025, www.twobirds.com/en/insights/2025/landmark-ruling-of-the-munich-regional-court-(gema-v-openai)-on-copyright-and-ai-training.
[2] Hembt, 2025.
[3] Kuschel, Linda, and Darius Rostam. “If It Looks Like a Duck: On the Munich Regional Court’s Ruling in GEMA v Open AI.” Verfassungsblog, 19 Nov. 2025, www.verfassungsblog.de/munich-regional-courts-ruling-in-gema-v-open-ai/.
[4] Scannell, Barry, and Leo Moore. “GEMA v OpenAI – Europe’s Direction on AI Infringement?” William Fry, 12 Nov. 2025, www.williamfry.com/knowledge/gema-v-openai-europes-direction-on-ai-infringement.
[5] Hembt, Dr. Simon, Dr. Niels Lutzhöft, and Toby Bond. “Long-Awaited German Judgment by the District Court of Hamburg (Kneschke v. LAION) on the Text and Data Mining Exception(s).” Bird & Bird, 10 Jan. 2024. https://www.twobirds.com/en/insights/2024/germany/long-awaited-german-judgment-by-the-district-court-of-hamburg-kneschke-v-laion.
[6] CEULI. “Navigating the Copyright Minefield: Legal Challenges of AI Training and Content Use.” 13 Jan. 2025. https://ceuli.com/navigating-the-copyright-minefield-legal-challenges-of-ai-training-and-content-use/#_ftnref32.
[7] Hembt, Lutzhöft, and Bond, 2024.
[8] Kuschel and Rostam, 2025.
[9] Kuschel and Rostam, 2025.
[10] Hembt, 2025.
[11] Scannell and Moore, 2025.
[12] Kalhor‑Witzel, Ronak. “Germany Delivers Landmark Copyright Ruling against OpenAI: What It Means for AI and IP.” Inside Tech Law, 17 Nov. 2025, www.insidetechlaw.com/blog/2025/11/germany-delivers-landmark-copyright-ruling-against-openai-what-it-means-for-ai-and-ip
[13] Hembt, 2025.
[14] Scannell and Moore, 2025.
[15] Kalhor‑Witzel, 2025.
[16] Kuschel and Rostam, 2025.
[17] Scannell and Moore, 2025.
[18] Kalhor‑Witzel, 2025.
[19] Hembt, 2025.
[20] Kuschel and Rostam, 2025.
[21] Scannell and Moore, 2025.
[22] Kalhor‑Witzel, 2025.
[23] Hembt, 2025.
[24] Hembt, 2025.
[25] Scannell and Moore, 2025.
[26] Kalhor‑Witzel, 2025.
[27] Kuschel and Rostam, 2025.
[28] Scannell and Moore, 2025.
[29] Kuschel and Rostam, 2025.
[30] On 3 April 2025, the Budapest Regional Court asked the CJEU to clarify in Like Company v Google Ireland (C-250/25) whether Google’s Gemini chatbot unlawfully reproduces protected press content. See: CJEU, Summary of the request for a preliminary ruling, Case C-250/25, Like Company v Google Ireland Limited, lodged 3 April 2025 (decision to refer dated 10 March 2025), under Article 98(1) of the Rules of Procedure. Working document available: https://curia.europa.eu/juris/showPdf.jsf?docid=300681&doclang=EN.
[31] Kuschel and Rostam, 2025.
