A Munich court has found that OpenAI violated German copyright and sided with music rights society GEMA in its closely watched lawsuit over the use of protected song lyrics for training large language models. The decision is for damages (the amount of which has not been disclosed) and represents a significant milestone in the fast-paced development of generative AI as it encounters IP law in Europe.
OpenAI said it disagrees with the ruling and was exploring its options, noting that the decision relates to a specific set of lyrics and would not impact daily use of its tools by individuals and businesses in Germany. GEMA celebrated the decision as a precedent that means AI developers must get permission when ingesting creative works.
What the Munich court ruling means for copyright
The question was whether both OpenAI’s model training and outputs have improperly utilized copyrighted song lyrics held by GEMA members without permission. The court did not publicly release the full reasoning behind its ruling, but the decision makes clear that training on and transferring copyrighted lyrics can be considered infringement under German law when it is done without a license or statutory exception.
The regional courts in Munich, Germany, are already well known for hearing high-profile intellectual property disputes, the outcomes of which often echo across the continent. In doing so, the court seems to grapple with a pair of central issues that are edging their way into AI cases: whether it is permissible to ingest copyrighted works at scale (the decision barely touches on this), and the danger that models will spit out protected text verbatim or near-verbatim because they have memorized it.
OpenAI said the disagreement was limited in nature. GEMA called it a fundamental question of principle: whether creative works can be fed into commercial AI systems without permission. The court’s damages order, while not yet reduced to a sum certain, represents significant exposure both from legal and financial perspectives for training models that are not definitively rights-cleared.
Why this case matters for the training of AI in Europe
There are exceptions (such as Text and Data Mining) in European copyright law, but there is a carve-out for commercial use if rights owners want to reserve their rights. Germany has in the Urheberrechtsgesetz (in particular Section 44b) provided that rights holders can opt out of text and data mining for commercial purposes. But if those reservations exist, developers need licenses to mine the works. Those rights are increasingly being enforced by music publishers and collecting societies.
GEMA serves 94,000 members and is one of the largest societies for authors of musical works in the world. Its licensing revenues annually top more than €1 billion, driving home what is at stake economically for creators as their works are fed into AI systems. The ruling strengthens the position held by European rights holders demanding payment or technical controls over how their catalogs are used to train artificial intelligence.
The decision also complements an EU-imposed need for foundation models to divulge information pertaining to training data sources. Transparency requirements and a well-defined path for rights reservation further incentivise AI developers to deploy provenance tracking, licensing or opt-out mechanisms.
Industry and legal reactions to the Munich court ruling
The result was spun by the chief executive of GEMA, Tobias Holzmüller, as a safety barrier for creative livelihoods — insisting that “the internet is not a self-service store” and that operators of AI tools must respect copyright. That position is part of a larger push by writers, publishers and labels to gain compensation and control in the age of generative AI.
OpenAI argued that the decision was circumscribed and said it may challenge it. The company has since been trying to negotiate licensing deals with content owners in a variety of other industries, in what would seem to indicate the industry’s shift toward negotiated access to data instead of blanket scraping. Lawsuits, meanwhile, have pushed the boundaries of model training: News publishers have accused companies of misusing articles in order to train models using copyrighted text; music firms have filed actions around reproduction of lyrics. In another development underscoring the size of the exposure, Anthropic settled a class action by authors over training on books for an amount in excess of $1.5 billion, according to public statements by those who were parties to the settlement.
What changes for developers and users after the ruling
For the AI builders in Europe, the message is clear: build strong rights management. That includes obtaining licenses when rights are reserved, respecting machine-readable opt-outs, documenting data lineage and deploying safeguards to ensure that models don’t spit out protected lyrics or text. Tricks such as de-duplication, dataset filtering, output watermark verification and fetch controls are now table stakes.
You shouldn’t expect there to be immediate service disruptions that you can see, but if it’s possible for content filters around lyrics and other works that are not yours to be tightened, they will be. Look for more licensing announcements, increased transparency into sources of training data and resources, and enhanced tools to enable creators to register reservations. The Munich decision is a spur to the more general trend that generative AI development is tacitly becoming subject to the same licensing logic which oversees streaming, broadcasting and online platforms.
The bottom line is all too clear: in the largest economy of the EU, courts are prepared to recognize unlicensed ingestion of protected works as a cause of action. For OpenAI and its peers, the road ahead resembles less an engineering hack than a licensing and regulation playbook.
