A German court has pushed one of the central questions in AI copyright into sharper focus: if a model can reliably reproduce protected lyrics, is the work effectively present inside the model itself?
The Munich Regional Court has mostly sided with GEMA, Germany's main music rights group, in its case against OpenAI. The decision grants requests for an injunction, disclosure, and damages, though the ruling isn't final yet.
What the Munich court decided
GEMA sued OpenAI late last year, alleging that ChatGPT reproduced copyrighted song lyrics without a license or payment to the writers. The case concerns song lyrics by nine well-known German writers, including popular songs that are widely recognized in Germany.
GEMA's argument was not simply that ChatGPT produced similar language. It said OpenAI's models had memorized the lyrics and could generate them almost word-for-word when prompted.
OpenAI disputed that framing. The company argued that its language models do not store specific training material, but learn general patterns from data. It also said any copyright questions should be handled under text and data mining rules.
The Munich court took a different view. It found that the disputed lyrics are reliably contained in the "4" and "4o" language models. After comparing the original lyrics with the model output, the court said the degree of similarity could not be explained as chance, given the length and complexity of the songs, according to its press release.
Why model parameters matter
The central issue is whether a protected work can count as reproduced when it is represented inside a model's parameters. For the Munich court, the answer can be yes.
The court treated the embedded lyrics as copyright-relevant reproduction. Its reasoning was that the lyrics are embodied in the model itself, even if they appear only as probability values rather than as a conventional stored file.
That matters because AI companies often separate training from output. They may argue that a model learns statistical relationships, not stored copies of individual works. The Munich decision narrows that distinction when memorization is shown and the protected material can be extracted through prompts.
The court also cited the EU directive on reproductions, which covers works "by any means and in any form, in whole or in part." In the court's view, that broad language was enough to include the way lyrics can be represented in model parameters.
Where text and data mining did not help
Text and data mining, often shortened to TDM, is a key part of the debate. In general, companies building large language models can make copies for analysis under TDM rules.
But the Munich court drew a line between making copies to create a training dataset and having a work embedded in the finished model. It said TDM covers copies needed to assemble the dataset. It does not, in the court's view, cover a copyrighted work becoming part of the model itself.
That distinction is important for AI copyright cases because it shifts attention from the act of collecting and processing training data to the state of the trained model. If a model can be shown to contain and reproduce protected material, the court's reasoning suggests the legal problem does not end once training is complete.
The ruling also addressed ChatGPT's responses. The court found that the chatbot's outputs amounted to unauthorized reproductions and public distribution of the lyrics because recognizable elements of the songs appeared in the generated text.
Why OpenAI was held responsible
The court placed responsibility on OpenAI as the operator, not on the users who entered prompts. That part of the decision is significant because it focuses on control over the system.
According to the court's statement, OpenAI runs the models, selects the training data, trains the model, and controls both the model's structure and what gets memorized. The court said this gives OpenAI decisive influence over what the model produces.
In practical terms, the court did not treat the chatbot as a neutral tool whose output is solely directed by users. It looked at the upstream choices behind the model and connected those choices to the downstream generation of protected lyrics.
For AI companies, that reasoning raises the stakes around memorization. If the operator controls the data pipeline, the model design, and the trained system, the operator may also face responsibility when copyrighted material appears in outputs.
A widening split with the UK
The Munich ruling stands in direct contrast to a recent UK decision involving Getty Images and Stability AI. The UK's High Court threw out Getty Images' claim against Stability AI, ruling that an AI model is not an "infringing copy" because its weights do not actually contain or reproduce copyrighted works.
That creates a stark difference between two courts within days. Munich says that defining a work in model parameters can be enough to count as reproduction. London says model weights contain learned patterns and features, not copyrighted works.
The result is a fragmented legal landscape for AI and copyright. In Germany, memorization can support injunctions, damages, and information claims. In the UK, model weights are not recognized as infringing copies at all.
The case is not final, so it is not the last word on OpenAI, GEMA, or copyright in AI models. But it does show how unsettled the rules remain. The same technical question, whether model weights or parameters contain protected works, is now receiving sharply different answers in different courts.