In November 2025, the Munich Regional Court delivered what is widely regarded as a landmark ruling: GEMA sued OpenAI, claiming the U.S.-based AI firm had used copyrighted German song lyrics to train its models and then reproduced them in its chatbot outputs. The Court sided with GEMA, stating that large language models (LLMs) can store and reproduce protected content and that unlicensed use of those works constitutes copyright infringement under German and EU law. This decision marks one of Europe’s first major judicial pronouncements on how generative AI technologies must respect traditional intellectual-property rules.
The Origins of the Case: Why GEMA Took Action
GEMA, Germany’s leading music-rights society representing over 100,000 composers, lyricists and publishers, noticed that certain musical works from its repertoire were being reproduced — verbatim or with only slight modifications — via ChatGPT prompts. According to press reports, some tunes by major German artists such as Herbert Grönemeyer (“Männer”) and Helene Fischer (“Atemlos durch die Nacht”) were among the list of nine songs identified in the lawsuit.
GEMA argued that OpenAI had trained its models on these protected lyrics without obtaining a licence, and that the chatbot’s ability to reproduce them pointed to unlawful copying. The society claimed that it had offered a licensing framework for AI developers as early as 2024, but OpenAI rejected it. By pursuing court action, GEMA sought not only damages but a clear statement that content creators’ rights cannot be bypassed by AI training practices.
The Court’s Decision: Key Legal Findings
Training = Reproduction
At the heart of the ruling was the court’s conclusion that training an AI model on copyrighted lyrics counts as a reproduction under Germany’s Urheberrechtsgesetz (UrhG) and the EU InfoSoc Directive (2001/29/EC). The judges held that even if the lyrics are not stored as human‐readable text but rather embedded in the model’s parameters (“weights”), they can still be reproduced upon user prompt. Thus, the fact that ChatGPT could output large portions of those songs signalled the presence of “fixed” copies inside the model.
Outputs = New Act of Infringement
The Court further ruled that when ChatGPT provided users with the protected lyrics, that amounted to a separate act of reproduction and communication to the public. Under UrhG § 19a (and InfoSoc Art. 3), providing access to the work is an exclusive right of the creator. The court assigned responsibility to OpenAI, not the end-user, because the provider chose the training data, designed the model and made output available.
TDM Exception Narrowly Interpreted
OpenAI had argued that its training activities fell under the Text and Data Mining (TDM) exception (Germany: § 44b UrhG; EU Directive 2019/790). The Munich court rejected this defence, saying the TDM exception covers analysis of content (e.g., linguistic features), not the embedding and later reproduction of entire expressive works. Because the songs were “memorised” and could be reproduced, the activity exceeded mere mining.
No Implied Consent or Fair Use
Unlike in some other jurisdictions, German law does not recognise a broad “fair use” or implied consent for training AI. The Court emphasised that creators do not automatically permit their works for any new use, especially when it involves deep-learning systems and commercial deployment. If data is used beyond what the rights-holder expected, a licence is required.
Why This Ruling Matters for AI Developers
Licensing Risk Is Real
For AI companies operating in Europe, this decision raises the risk bar significantly. Training models on copyrighted content — especially short, expressive works like song lyrics, poems or code — now demands clearer legal scrutiny. Blanket assumptions about scraped data, or reliance on “public domain” ambiguities, may no longer suffice.
Technical & Organisational Impact
Providers may now need to adopt robust data governance, including:
→Deduplication and de-sensitisation of training data
→Filters to prevent verbatim output of protected works
→Licensing frameworks or partnerships with rights-holders
Legal analysts anticipate that AI firms will revise training methodologies to reduce memorisation risks, particularly for content easily reproduced.
Strategic Positioning in Europe
Given that Europe’s regulators (and upcoming AI regulations such as the EU AI Act) emphasise transparency and rights-holder protections, this ruling aligns with a wider trend. AI companies wishing to scale in Europe must now consider rights-based compliance as part of their business model, not just engineering.
Creators’ Leverage Increases
For music artists, authors and publishers, the verdict is a major win. It sends a clear message: if your work fuels AI training, you have a claim — not only for output but for the training stage itself. Collecting societies may now pursue dual enforcement: training-stage use + output-stage dissemination.
How Europe’s Approach Compares to the U.S. and Other Jurisdictions
United States: Fair Use Focus
In the U.S., copyright law is more flexible, thanks to the concept of fair use. Several lawsuits (e.g., authors vs. AI firms) are ongoing, but American courts tend to weigh factors such as transformation, market harm and substitution rather than focusing solely on “copying”. Some U.S. judges have found that training an AI model might be a fair use if it involves abstract learning and the output does not reproduce substantial, recognisable parts of copyrighted content.
United Kingdom: Technical Distinctions
In a contemporaneous case, the UK High Court (in Getty v Stability AI) held that a diffusion model did not constitute an infringing copy, because the model did not store or deliver identifiable copyrighted images. The UK reasoning differs markedly from the German court’s view of “memorisation” as a copy.
Why Europe May Diverge
German and EU law treat copyright as an economic property right; reproduction is defined broadly (“in any form and by any means” under InfoSoc Art 2). The Munich ruling reflects that doctrine — viewing embedded model weights as functionally equivalent to storage of the works. U.S. and U.K. law are more inclined to focus on output risk, market impact, and whether the use is transformative. The divergence means AI firms face jurisdiction-specific risk profiles — what’s permissible in one geography may be infringing in another.
Voices From the Field: Creators, Technologists, Regulators
Creators: GEMA’s CEO, Tobias Holzmüller, celebrated the judgment: “The internet is not a self-service store, and human creative achievements are not free templates.”
Many songwriters welcome the ruling as long overdue — they argue their work is treated as raw material by AI firms without compensation or recognition.
Technologists: On the AI side, OpenAI stated that the case covers only a limited set of lyrics and that they are considering an appeal, while emphasising the importance of large data sets for innovation. Some in the AI community warn that overly rigid rights regimes might hamper Europe’s competitiveness in AI development.
Regulators: European policymakers are watching closely. The European Commission has already stressed that copyright must be respected and that training data must be documented. The upcoming AI Act may require model providers to demonstrate compliance with rights-holders and licensing frameworks. This ruling strengthens the regulatory narrative that AI must operate within existing legal frameworks, not beside them.
What This Means for the Future of AI and Copyright
A New Business Model for AI Training
We’re likely to see licensing arrangements emerge, where AI firms pay rights-holders for access to content. Similar to streaming-music licences, we might get “AI-training licences” for songs, books, images and other high-value content.
Technological Adjustments
AI developers might shift to public-domain datasets, synthetic data or licensed corpora, and enhance safeguards for output filtering. Training strategies may favour abstraction over memorisation, improving legal safety while preserving model performance.
Legal Clarification and Reform
This ruling could prompt further court decisions and possibly an appeal to the Court of Justice of the European Union (CJEU). It may also influence how the EU interprets the TDM exception and copyright scope for AI training.
Creativity & Culture Still Matters
Beyond law and technology, this case raises fundamental questions: Who owns the building blocks of culture when machines learn from them? Can AI truly be ‘creative’ if its training data is itself creative work? The ruling affirms that human creativity deserves protection, even in the age of machines.
Final Thoughts
The GEMA vs. OpenAI decision is a watershed moment. By holding a leading AI provider accountable for unlicensed use of copyrighted musical works, a German court sent a clear signal: AI innovation cannot sideline creator rights. For rights-holders, it provides new legal footing. For AI companies, it adds another compliance dimension. And for society, it reinforces that technological progress and cultural respect must go hand in hand. As generative AI continues to redefine how we create and consume content, this case will likely be studied for years to come — not just for its immediate outcome, but for what it means for the balance between machines, creators and law.
Stay curious, stay informed, and let´s keep exploring the fascinating world of AI together.
This post was written with the help of different AI tools.
