In a landmark decision, the 42nd Civil Chamber of the Munich I Regional Court has ruled in favor of GEMA, Germany’s principal music-rights society, finding that OpenAI’s language models infringed copyright by memorizing and reproducing song lyrics.
The ruling, handed down in Case No. 42 O 14139/24, addresses nine well-known German songs, including Kristina Bach’s “Atemlos” and Rolf Zuckowski’s “Wie schön, dass du geboren bist.” While the court dismissed claims based on alleged violations of the general right of personality due to incorrect attribution of altered lyrics, it essentially granted GEMA’s requests for injunctive relief, disclosure, and damages.
According to the court, the song lyrics in question were reproducibly contained in OpenAI’s language models and were later reflected in the outputs produced by the company’s chatbots. The court emphasized that this phenomenon, known in AI research as “memorization,” occurs when a model not only extracts patterns from training data but also incorporates entire works into its parameters.
By comparing the lyrics used in training with the outputs generated in response to simple user prompts, the court concluded that the reproduction was not coincidental. This memorization, the court explained, constitutes a reproduction under German copyright law and Article 2 of the EU InfoSoc Directive, even if the work is encoded within the probabilistic parameters of a model rather than stored in a conventional format.
The ruling explicitly rejected OpenAI’s argument that any potential infringements were covered by limitations for text and data mining under Section 44b of the German Copyright Act. The court noted that the legal justification for text and data mining applies only to preparatory acts, such as copying works into RAM for analysis, and not to the permanent incorporation of copyrighted material into a model that can subsequently generate the work in outputs. The court also dismissed claims that the reproductions were incidental under Section 57 of the Copyright Act, finding that the lyrics were distinct and protected works and could not be considered trivial alongside the broader training dataset.
Importantly, the court determined that OpenAI, as the operator of the language models, is responsible for these infringements. It rejected the company’s contention that users generating outputs were solely liable, stressing that the architecture of the models and the selection and memorization of training data fell entirely within OpenAI’s sphere of influence. Consequently, the outputs produced in response to simple prompts were also considered infringements, and the company could not rely on any limitation or defense to avoid liability.
For enterprises that develop or deploy AI systems, the ruling carries broad implications—particularly when compared to the recent ruling in the UK High Court in Getty Images v. Stability AI. Companies relying on third-party AI models trained on unlicensed or poorly documented data may now face legal exposure if outputs inadvertently reproduce copyrighted content. Organizations must assess data provenance, consider contractual protections such as warranties and indemnities from vendors, and monitor outputs for potential copyright violations. Enterprises may increasingly prefer models built from fully licensed or proprietary datasets, or adopt retrieval-augmented generation that draws on internal data rather than large, uncurated public corpora.
The ruling also signals that copyright compliance should be a central consideration in AI strategy, particularly for multinational enterprises operating in jurisdictions with strong IP enforcement. While the case focuses on German law, it could influence broader EU legal interpretations and shape regulatory expectations for AI systems worldwide. Enterprises should anticipate potential licensing obligations, increased operational costs, and reputational risk if AI outputs inadvertently infringe rights.
Although uncertainties remain—such as how this precedent may extend to books, images, or software code, and how future licensing frameworks for AI training data might evolve—the Munich court’s decision is a clear signal that generative AI cannot be deployed without careful consideration of copyright. For enterprises, the case underscores the need to embed rights compliance into AI governance, vendor management, and operational oversight, ensuring that innovative AI applications do not come at the expense of creators’ legitimate interests.