GEMA vs. OpenAI | AI memorisation is a reproduction relevant to copyright law, and the TDM exception does not help in LLM training, Munich I Regional Court holds

In its judgment of 11 November 2025 (42 O 14139/24), the Munich I Regional Court (Germany) issued a widely noted precedent on the copyright assessment of AI training and outputs under German and EU law. According to its press release¹, the Court found that large language models (LLMs) could memorise copyright-protected texts and that this fixation within the model already constituted reproduction under copyright law. If the LLM reproduced essential and original elements of such texts in response to simple prompts, that was a further act of reproduction, according to the Court. Providing the output to users further constituted making the work available to the public and the text and data mining (TDM) exception did not apply in this case, the Court found.

Thus, the Court granted injunctive relief and information claims and found liability for damages, the quantum yet to be determined. However, claims based on personality rights for incorrect attribution were dismissed. The judgment is not yet final.

What was the case about?

The Court dealt with a lawsuit brought by GEMA against OpenAI, provider of ChatGPT, whose versions 4 and 4o allegedly had "fixed" nine well-known German song lyrics in the LLM in a way that, in response to simple prompts, they were output largely true to the original. OpenAI alleged that the ChatGPT did not store specific training data but only statistical knowledge in its parameters and claimed that any infringements were caused by user prompts, whereas the training was covered by the TDM exception.

Memorisation as reproduction, according to the Court

The key issue for the Court to decide was whether so‑called memorisation constituted reproduction under Sec. 16 of the German Copyright Act (Urheberrechtsgesetz, UrhG) and Art. 2 InfoSoc Directive (2001/29/EC). Reproduction in this sense covers copies “in any form and by any means” and does not require direct perceptibility. The Court held it to be sufficient if protected texts were encoded in model parameters so they could be largely faithful to the original when extracted via simple prompts. The fact that this encoding takes the form of probability weights did not preclude classification as a copy under copyright law, the Court underlined. Relying on research into the extractability of training data from large language models, the Court ruled out coincidence as an explanation for the outputs observed.

Outputs and responsibility – user or provider?

The Court also considered the output of copyright-relevant content to be an independent act of use under copyright law: Insofar as original recognisable elements of certain song lyrics appeared in the output, this constituted further reproduction under Sec. 16 UrhG. At the same time, making this output content available on demand to a potentially indefinite number of users signified making it publicly available within the meaning of Sec. 19a UrhG and Art. 3 InfoSoc Directive, the Court held. The Court put primary responsibility for this on the provider, which controlled architecture, data selection and training of the LLM and therefore memorization and its risks; simple user prompts did not shift this responsibility to end users.

No limitations or consent "by virtue of custom" applied

The Court clearly rejected OpenAI’s reliance on the TDM exception, Sec. 44b UrhG to permit only preparatory copies needed for analytical purposes. Once the LLM produced reproducible works, the exploitation interests of right holders were affected and the scope of the TDM exception was exceeded, according to the Court, which also refused any extended or analogous application on the basis of the provision’s wording, structure, and purpose. Moreover, Sec. 57 UrhG on insignificant accessories did not apply, the Court held, because there were no protected main work and a training dataset itself was not a work. Research exceptions for TDM under Sec. 60d UrhG and Art. 3 DSM Directive ((EU) 2019/790) were not relevant. The Court also did not assume consent by custom because it did not consider training generative AI to be a use that right holders had to expect normally.

Counterpart to the London High Court’s ruling in Getty Images v Stability AI?

In an international comparison, a clear contrast to the decision of 4 November 2025 which the London High Court had delivered in Getty Images v Stability AI only one week earlier stands out, at least on the outcome.

The High Court dismissed Getty’s secondary copyright infringement claim because it did not consider Stable Diffusion, Stability AI’s model at stake, to be an "infringing copy". The Munich Court, on the other hand, expressly assumed that an AI model (in its case, ChatGPT) might "contain" copies of the training material that were relevant for copyright purposes if the protected works were reproducibly defined. This is a fundamental different understanding of the technical facts of generative AI systems and the implications under copyright law.

However, the English High Court did not decide on Getty’s core allegation that Stability AI had used millions of copyright-protected Getty photos to train Stable Diffusion as Getty could not establish that the training took place in the UK to make out such primary copyright infringement claim. Hence it remains open, how the High Court would decide on AI training with (allegedly) copyrighted content from third parties. In this respect, it remains to be seen whether there is also such a divergence between German and EU law on the one hand and UK law on the other hand.

Outlook for the future:

In Germany, the Munich Court’s ruling noticeably shifts risk toward AI providers and increases licensing pressure as well as technical and organisational obligations. Providers should prioritise deduplication, regularisation, targeted anti-memorisation techniques and robust prompt and output filters, particularly for short works such as song lyrics and poems, if they want to avoid liability risks under copyright law. Rights holders and collecting societies, in contrast, may have gained a dual enforcement lever because they might base claims on training-stage memorisation as well as on infringing outputs.

However, unless the dispute is settled between GEMA and OpenAI, the legal community may certainly expect an appeal against this decision and it is unlikely that the Regional Court’s judgment will be the last word in this area, which seems fundamental for the balance of rights between copyright owners and AI providers.