Landmark ruling of the Munich Regional Court (GEMA v OpenAI) on copyright and AI training

What was the case about?

Germany’s music collecting society, GEMA brought an action before the Munich I Regional Court (case no. 42 O 14139/24) against the US-based AI developer OpenAI, alleging that the company had used protected song lyrics when training its 4 and 4o GPT large-language models which power ChatGPT without obtaining a licence.

The dispute concerned the lyrics of nine well-known German songs – including “Atemlos” (Kristina Bach), “Männer” (Herbert Grönemeyer) and “Über den Wolken” (Reinhard Mey). 

GEMA argued that the lyrics had been memorised in the GPT model parameters and could be reproduced almost verbatim, amounting to an unauthorised reproduction under Sections 15, 16 and 19a of the German Copyright Act (Urheberrechtsgesetz, UrhG).

OpenAI maintained that its models do not store or copy specific training data but instead reflect statistical correlations learned from the dataset as a whole. Any text generated would, it argued, result from user prompts beyond its control. OpenAI also relied on the EU text- and data-mining (“TDM”) exceptions in Articles 3 and 4 of Directive (EU) 2019/790 (the “CDSM Directive”, implemented in Sections 60d and 44b UrhG), arguing that such uses were, in any event, covered by those exceptions.

What did the Munich court decide and why does it matter?

In a landmark first-instance ruling, the 42nd Civil Chamber of the Munich Regional Court largely upheld GEMA’s claims for injunctive relief, disclosure and damages, while dismissing a secondary claim concerning alleged personality-right violations.

The court found that simple user prompts led ChatGPT to reproduce substantial parts of the original lyrics almost verbatim. The court noted that hallucinations were present in some outputs but concluded that these deviations did not affect the recognisability of the original song lyrics in their peculiarity, as the memorised content itself remained consistent and the variance primarily concerned interchangeable introductory and concluding texts rather than the core lyrical content. Given the complexity and length of the texts, the court concluded that coincidence was excluded. Both parties agreed that the song lyrics had been used for training but disagreed on whether this constituted “copying.”

Relying on findings from scientific literature in information technology research, the court accepted GEMA’s argument that training data can become embedded in model weights and remain retrievable – a phenomenon referred by GEMA as memorisation. Observing that the works were reproducible through very simple prompts such as "How is the text of [song title]" and "What is the refrain of [song title]", the court considered the statutory notion of fixation under EU law to be satisfied. In doing so it rejected OpenAI’s argument that it was necessary for GEMA to identify specifically, definable data in the model. By drawing an analogy with the lossy compression method used to encode MP3 files, the court concluded it was sufficient that a model could generate statistically probable token sequences that recognisably reproduce the song lyrics on the basis of the statistical information derived during training. 

Based on these findings, the court held that the presence of the lyrics in the model parameters amounted to fixation, and that their reproduction in ChatGPT outputs represented additional acts of reproduction and communication to the public. Referring to Sections 15, 16 and 19a UrhG and Articles 2 and 3 of Directive 2001/29/EC (the “InfoSoc Directive”), the court found that the memorisation of lyrics in the models “4” and “4o” of ChatGPT constituted unauthorised reproduction and communication to the public.

Which legal questions did the court address?

• Did the court find that AI training constitutes “reproduction” under German copyright law?

Yes. Following Article 2 InfoSoc Directive, the court held that a reproduction exists “in any form and by any means.” Even a fixation through numerical probability values qualifies, as long as the work can later be perceived through technical means. The court considered the model parameters to embody the protected expression.

• Why did the court not apply the TDM exceptions?

The court confirmed that training large language models will generally fall within the scope of application of the text and data mining barriers, with the German legislator explicitly listing “machine learning as a basic technology for artificial intelligence” within the scope of application of Section 44b UrhG. However, the court found that the reproduction of the disputed song lyrics in the models does not constitute text and data mining, as text and data mining aims at the evaluation of information such as abstract syntactic regulations, common terms and semantic relationships, whereas the memorisation of the song lyrics at issue exceeds such an evaluation and is therefore not mere text and data mining

• Did the court consider any other exemptions or implied consents by the authors?

No. The court stated that training AI models is not an ordinary or expected use of a work to which authors have implicitly consented. The acts were therefore unlicensed. Furthermore, the court found that the use was not justified by quotation, parody or similar limitations to copyright.

• Who did the court find to be responsible for the AI outputs?

The court determined that responsibility lies with OpenAI. The company selected the training data, built and operated the system, and determined its architecture. User prompts merely trigger the model’s internal processes and do not create independent liability.

• How should an AI developer comply with this judgment according to the court?

The court acknowledged that compliance may pose technical challenges, since training data cannot easily be removed once embedded in a model. Nevertheless, it placed the obligation on providers to prevent further infringements. In practice, this may require layered mitigation measures – such as internal guidelines restricting outputs, filters to block reproductions, and licensing or retraining strategies for future versions.

What comes next?

OpenAI has announced plans to appeal. The judgment is not yet final and may be reviewed by the Munich Higher Regional Court. A reference to the Court of Justice of the European Union also remains possible.

Meanwhile, GEMA continues separate proceedings against Suno AI concerning AI-generated music.