On November 11, 2025, the Munich Regional Court I (Landgericht München I) issued a first-of-its-kind decision in Europe, holding OpenAI liable for copyright infringement. The case was brought by GEMA, Germany’s collective rights management organization representing over 100,000 composers, lyricists, and music publishers. At the heart of the dispute: the use of copyrighted song lyrics to train the GPT-4 and GPT-4o language models powering ChatGPT.
The ruling reignites the debate over how far AI companies can go in using protected works to train their systems — and carries significant implications for rights holders worldwide, including those with interests in Brazil.
The Case
In case no. 42 O 14139/24, presided over by Judge Elke Schwager, GEMA demonstrated that ChatGPT could reproduce, nearly verbatim, the lyrics of nine German songs protected by copyright.
Simple prompts such as “What are the lyrics to [title]?” or “What is the chorus of [title]?” were enough to generate extensive, recognizable passages from the original works. Given the complexity and length of the compositions, the court ruled out coincidence as an explanation.
According to the court, the fact that basic prompts yielded such faithful reproductions indicated that the lyrics had been effectively memorized during model training.
The Arguments
GEMA argued that OpenAI systematically used its repertoire to train its language models without authorization or licensing. The organization contended that both the incorporation of lyrics into the model’s parameters and their reproduction in ChatGPT’s outputs constituted unauthorized acts of reproduction and communication to the public, violating Sections 15, 16, and 19a of the German Copyright Act (Urheberrechtsgesetz – UrhG), as well as Articles 2 and 3 of the EU InfoSoc Directive (2001/29/EC).
OpenAI countered that its models do not store or copy specific training data, but merely reflect statistical correlations learned from the dataset as a whole. The company also sought to shift responsibility to users, arguing that outputs result from user-provided prompts.
The Ruling
The Munich court rejected OpenAI’s defense entirely. Drawing on computer science literature, the court accepted GEMA’s argument that training data can become embedded in model weights and remain retrievable—a phenomenon known as “memorization.”
Using an analogy to lossy compression (as in MP3 files), the court held that fixation of a work does not require literal storage; it is sufficient that the model can generate statistically probable token sequences that recognizably reproduce the original content. Accordingly, the presence of lyrics in the model’s parameters was deemed a form of fixation, and their reproduction in ChatGPT’s responses an additional act of reproduction and communication to the public.
The court also declined to apply the text and data mining (TDM) exceptions under Articles 3 and 4 of the EU Digital Single Market Directive (2019/790), transposed into Sections 60d and 44b of the German Copyright Act. These provisions permit automated extraction of information from texts, images, and other digital content for analysis purposes. However, the court found that memorizing song lyrics exceeds this purpose: it is not mere extraction for analysis, but incorporation of protected works in a way that enables their full reproduction.
OpenAI was ordered to pay damages (amount undisclosed) and cease using the works. The decision is not yet final and may be appealed to the Munich Higher Regional Court (OLG München), the German Federal Court of Justice (Bundesgerichtshof), and potentially the Court of Justice of the European Union.
International Contrast
The German ruling gains further significance when compared to another recent judgment. Just one week earlier, on November 4, 2025, the High Court of London examined a similar case: Getty Images sued Stability AI, developer of the Stable Diffusion image generation model, alleging that the system had been trained on millions of protected photographs without authorization. The UK court reached the opposite conclusion, holding that an AI model does not itself constitute an infringing copy of the works used in training—even if protected images were used to “teach” the system, the resulting model is not, by itself, a copyright violation.
This divergence underscores the unsettled state of AI copyright law globally.
Implications for Brazil
The German decision offers important insights for Brazil’s legal framework. The Brazilian Copyright Act (Law No. 9,610/98), Article 29, requires prior express authorization from the author for any use of a work, including partial or full reproduction. Unlike the European Union, Brazil has no specific legal exceptions for text and data mining that might support the use of protected works in AI training.
Thus, the reproduction of protected works in AI system outputs—as occurred in GEMA v. OpenAI—would, in principle, also constitute copyright infringement under Brazilian law, regardless of whether the work was actually stored or merely “learned” by the model.
An open question remains: does the training of AI models on protected works, independent of the output, constitute unauthorized reproduction under Brazilian law? The absence of specific regulation leaves room for divergent interpretations and creates legal uncertainty for developers, users, and rights holders alike.
The German case may serve as a reference for future legislative discussions and court decisions in Brazil, much as the DABUS case did. In that matter, patent applications were filed in several countries naming an AI system as the inventor. While some jurisdictions (such as South Africa) accepted this possibility, most patent offices—including the European Patent Office, the USPTO, and Brazil’s INPI—rejected such applications, holding that current law presupposes a natural person as inventor. DABUS influenced global debates on the limits of IP protection for AI-involved creations, and the German copyright ruling may have a similar ripple effect.
The Importance of Specialized Counsel
The Munich court’s decision marks a milestone in European jurisprudence on artificial intelligence and copyright. By recognizing that memorization of protected works in language models constitutes unauthorized reproduction—and by rejecting the text and data mining exceptions — the court signaled that creator protection must prevail even amid the technical particularities of AI systems.
Although the ruling may still be appealed, it sets a relevant precedent likely to influence courts and legislators worldwide. For composers, lyricists, publishers, and other rights holders, now is the time to assess how their works are being used by AI platforms and what protective measures can be adopted.
Given the technical and legal complexity at the intersection of intellectual property and artificial intelligence, specialized counsel is essential to adequately protect rights over creative works and navigate this rapidly evolving landscape.
Ricci Propriedade Intelectual closely monitors global developments in AI and copyright. Our team is available to advise artists, composers, publishers, and companies on protection strategies and monitoring of their works in digital environments—including matters involving Brazilian intellectual property law.
