Skip to content 
In a ruling issued this month, November 2025, the Regional Court of Munich (Germany) determined that OpenAI’s training and deployment of its large language model (LLM) system constituted a violation of copyright. The lawsuit was initiated by GEMA, the German collecting society representing music creators, on behalf of various authors and lyricists. The ruling signifies a significant milestone in the convergence of generative artificial intelligence (AI) and copyright law within the European Union.
This article examines the facts of the case, the court’s reasoning, and the implications for AI developers and rights-holders—who are engaging with AI technology and intellectual property.
Background to the copyright dispute
GEMA, representing approximately 100,000 composers, songwriters, and publishers in Germany, initiated legal proceedings against OpenAI in November 2024. The primary allegation was that OpenAI’s model was trained on copyrighted song lyrics without proper licensing or authorisation and that responses generated by the ChatGPT system (as well as other large language models operated by OpenAI) reproduced significant portions of those lyrics when prompted. Specifically, the case concerned nine German compositions within the repertoire of authors represented by GEMA. The assertion was that OpenAI’s system had memorised the lyrics and that users of the chatbot could prompt it to produce exact or nearly exact reproductions of the lyrics.
OpenAI justified its position on two fundamental points:
Its models do not retain or replicate training data verbatim; rather, they acquire statistical patterns and produce original output. Any output generated is user-produced (through prompts), and consequently, responsibility lies with the user rather than the operator of the model. In response, GEMA affirmed that the model indeed reproduced copyrighted works, that the training process was conducted systematically and without proper licensing, and that rights holders are entitled to licensing or compensation.
The Court’s Ruling
On 11 November 2025, the Regional Court of Munich I (42 O 14139/24) gave judgement to GEMA. Key findings and underlying reasoning are as follows:
Memorisation equals reproduction
The court found that the LLM’s limits did include reproductions of the protected works. Even if the model did not store them in the manner of a text file, the ability to reproduce the lyrics in the output upon prompt indicated that a reproduction had taken place. In the court’s opinion, copyright law does not necessitate an exact duplicate of the file; a reproduction within the model that is accessible to the public through the output is sufficient.
Training and output both infringe
The infringement was caused not only by the model’s output but also by the training process. The court determined that the taking over of the lyrics into the model parameters constitutes a reproduction under German and EU copyright law, specifically Section 16 of the UrhG and Article 2 of the InfoSoc Directive.
The TDM (Text and Data Mining) Exception Does Not Encompass This Application
OpenAI attempted to invoke the text-and-data-mining exception pursuant to § 44b UrhG (Germany) and Article 3 of the DSM Directive; however, the court dismissed this defence. It concluded that the exception applies solely to preparatory acts of reproduction (for analysis) and does not extend to storing protected works within model parameters for subsequent public dissemination.
Operator Responsibility
The court dismissed the assertion that accountability rested with the user prompt rather than the operator. Since OpenAI oversaw the model’s design, training, and deployment, it bore responsibility for the infringing acts.
Injunctions, Disclosure, and Damages
The court issued an injunction requiring OpenAI to provide an account and potentially compensate GEMA for damages. The exact quantum was not publicly disclosed at the time of reporting.
The Significance of This Decision
For professionals providing guidance on AI, intellectual property, and technology law, this judgement highlights several important considerations:
To date, the majority of litigation concerning AI training data and generative models has been centred on the United States. This German ruling constitutes one of the earliest significant European precedents concerning generative AI and copyright. The court’s perspective that “memorisation” within a model may comprise reproduction is especially significant.
Licensing Will Emerge as a Fundamental Concern – GET A LICENSE, YOU AI RASCALS, PAY FOR IT! [1]
Such is the arrogance of the AI developers; the ruling affirms that AI developers cannot rely solely on the extensive collection of copyrighted works without obtaining proper licensing. Indeed, GEMA has proposed licensing frameworks for the utilisation of instrumental works by AI, and the ruling compels developers to collaborate with rights holders.
Business Risks for AI Users and Providers
Although the decision was adverse to OpenAI, the underlying reasoning may have implications for other AI operators and providers of large language model services. Entities utilising or developing generative AI must assess their training datasets, licensing status, the risk of user prompts generating protected content, and the provisions for indemnity. Advisors serving corporate clients should identify and highlight potential liabilities.
Incremental Clarification of the Scope of TDM and Exception Regimes
The court’s restriction on the TDM exception clarifies that this exception limits permission for AI training that produces memorised and reproducible outputs. This implies that organisations cannot presume that textual and data mining alone absolves them of copyright exploitation rights.
Potential Broader Implications for the EU
Although this judgement pertains to a German national court, the fundamental interpretation of the InfoSoc Directive and DSM Directive indicates its relevance throughout the European Union. Other courts are likely to cite it, and we may anticipate referrals for preliminary rulings to the Court of Justice of the European Union in due course. Indeed, practitioners are already highlighting potential instances of divergence or convergence within case law.
Implications and Practical Considerations for Clients
For clients of Lawdit Solicitors in the areas of technology, media and communications, or corporate AI implementation, the following are essential practical considerations:
Perform Training Data Audits
Clients are advised to conduct an audit of the datasets employed for training or fine-tuning generative AI models. They should determine whether copyrighted materials (such as song lyrics, literary works, media transcripts, etc.) were incorporated without permission, and if so, whether the outputs could potentially reproduce such works when prompted.
Evaluate the Risk Associated with Model Output
Even in cases where training is lawful, the model may reproduce protected content through prompting. Clients should evaluate whether their model presents such risks and establish safeguards (such as filtering and response barring) to prevent the dissemination of copyrighted material.
Secure Licensing and Rights Acquisition
Where protected works have been utilised or will be utilised in training, a licence (or rights clearance) may be necessary. Clients should collaborate with collecting societies, rights-holders, or aggregators to negotiate AI-specific licensing agreements or reorganise training data to eliminate the necessity.
Examine Contracts and Indemnity Provisions
Clients offering generative AI services to third parties should ensure that contracts explicitly assign responsibility for outputs and should consider including indemnity provisions to address copyright risks. Providers may be required to establish compliance mechanisms, filtering processes, or remediation obligations.
Oversee Legal and Regulatory Developments
Given the novelty and swift development of generative AI and copyright legislation, clients should stay attentive to ongoing case law, regulatory guidance—including the upcoming EU AI Act—and amendments to copyright law at both national and EU levels.
Revise Internal Governance and AI Ethics Policies
Clients should incorporate copyright risk considerations into their AI governance frameworks, including training datasets, model design, output monitoring, procedures for addressing rights-holder complaints, and adherence to “fair remuneration” standards for creators.
Special Consideration for Music and Lyrics Use Cases
Use cases involving the generation of musical lyrics, assistance with songwriting, or music-based models are now particularly impacted, as this ruling pertains specifically to song lyrics. Clients operating within the music technology, entertainment, or audio industries should revaluate their exposure and licensing approaches.
Constraints and Future Directions
A few important caveats warrant mention:
The Munich decision remains pending finalisation; OpenAI has signalled its intention to pursue an appeal. The factual scope is limited— comprising only nine tracks within the GEMA repertoire. The ruling does not yet encompass all categories of protected works or geographic regions, although its reasoning may potentially be extended. The ruling pertained to German national law (UrhG) and the European Union copyright framework. In other jurisdictions, such as the United Kingdom and the United States, the position may vary significantly.
The court did not, in its public report, refer the matter to the Court of Justice of the European Union (CJEU) for a preliminary judgement, although commentators anticipated that issue would be raised. Despite these caveats, the decision remains a significant risk indicator that establishes a precedent, not a fully developed regime. Lawdit’s clients should therefore regard it as an urgent warning and a basis for strategic planning, rather than a secure and established refuge.
In conclusion
For the legal community and technology-driven enterprises in the United Kingdom (and worldwide), the Munich ruling in behalf of GEMA against OpenAI represents a major victory. It emphasises that operators of generative AI cannot presume copyright immunity solely due to automated or “pattern-based” training. The notion of “memorisation” within an AI model now entails legal implications under European copyright law.
For clients involved in developing, deploying, or licensing generative AI, the course of action is straightforward: conduct audits of training data and outputs; obtain rights or revise training methodologies; update policies and contractual agreements; and continuously monitor legal developments. For rights-holders and creators, the judgement indicates a viable enforcement pathway and serves as leverage to negotiate licensing agreements tailored to AI technologies. At Lawdit Solicitors, we provide guidance to clients across the full spectrum of intellectual property, technology, and artificial intelligence matters—whether you are an AI developer, a media enterprise licensing content, or a rights-holder aiming to safeguard your catalogue. The German ruling should prompt a reassessment of your AI strategy and intellectual property risk profile at this time.
[1] The view of the writer!
