GEMA v OpenAI: song lyrics in an AI model are copies

What the GEMA v. OpenAI ruling was about

„Atemlos“ by Kristina Bach or a birthday song by Rolf Zuckowski: anyone who asked ChatGPT for such lyrics received them largely word for word. GEMA, the collecting society that administers the rights of lyricists and composers, saw this as a copyright infringement and sued two companies in the OpenAI group. The Regional Court of Munich I largely found in its favour in a judgment of 11 November 2025 (42 O 14139/24). A language model that can reproduce song lyrics has reproduced them, and the exception for text and data mining does not cover this.

The 42nd Civil Chamber, which specialises in copyright, granted GEMA claims for injunctive relief, information and damages. It dismissed only one secondary point: an infringement of personality rights because the chatbot attributed altered lyrics to the actual authors. The judgment concerns nine sets of lyrics by well-known German authors. It is not final; OpenAI has appealed.

The defence: the model only learns

OpenAI raised three objections. First, the models did not store or copy any particular training data. The parameters, that is the numerical values set during training, merely reflected what the model had learned from the data set as a whole.

Second, the outputs only came about through user inputs, which is why the user, as the producer of the output, was responsible for it. Third, any interference was covered by the exception for text and data mining. This exception permits protected works to be analysed automatically in order to extract information and patterns from them.

Memorisation: the model knows the text

The court disagreed. It is known from computer science that training data can be contained in language models and can be retrieved again as output. The technical term is memorisation: during training, the model does not merely extract information from the data, it takes a text in its entirety into its parameters. The chamber established this by comparing the lyrics in the training data with the chatbot's outputs. With texts of this length and complexity, coincidence is ruled out as an explanation.

The court therefore reasoned backwards from the result. Anyone who recites a text almost verbatim on a simple request has stored it. The argument that one merely understood the poem and by no means learned it by heart is hardly convincing if all the verses then follow.

Storing in parameters is reproduction

In copyright law, reproduction means fixing a work in material form, for instance on paper or in a file. According to the judgment, fixing it in a model's parameters is also sufficient, as long as the text can be retrieved from them again. The fact that the work is held there only as probability values is irrelevant. European copyright law covers reproductions „by any means and in any form“.

Under the case law of the European Court of Justice, it is sufficient that the work can be made perceptible with technical aids. A language model is such an aid. Two further objections failed. The lyrics were not an incidental addition alongside the training data set as a whole, because that data set is not itself a protected work. And there is no implied consent on the part of the rightholders, because training AI models is not a customary use that an author must anticipate.

Text and data mining does not cover this

The key point of dispute was the text and data mining exception. According to the judgment, language models fall within its scope in principle. However, only the preparatory copies are covered: converting a work into another format, caching it in memory, compiling the data corpus. These copies serve analysis alone, do not affect the author's exploitation interests and are therefore free of remuneration.

Where, by contrast, the work itself is reproduced during training, that is no longer text and data mining. A technology- and innovation-friendly interpretation that would also exempt copies within the model is ruled out by the clear wording. Where copies are held in the model, exploitation of the work is lastingly impaired, and the risk of memorisation arises solely within the providers' sphere. Shifting it onto the rightholders would leave them without protection.

The provider is liable, not the user

The outputs themselves also infringe copyright according to the judgment. The chatbot reproduces the lyrics and makes them publicly available, because their characteristic elements remain recognisable throughout. The providers are responsible for this, since users merely entered simple prompts. OpenAI operates the models, selected the lyrics as training data and is responsible for the models' architecture and thus for the memorisation. The specific content of the answer is generated by the language model; whoever enters the prompt merely triggers the output.

What the judgment means for businesses

For companies that train or fine-tune their own models, the judgment shifts the allocation of risk. If protected texts remain retrievable in the finished model, a licence or technical safeguards against memorisation are needed. Anyone who does not know what their model has learned by heart should test it before others do. The comparison of training data and outputs serves as a template for this.

Anyone integrating a third-party model into their own products, for example into a customer-facing chatbot, will not find a ready answer in the judgment. The reasoning turns on control over architecture and training, and only the model provider has that. Whether the integrator nonetheless has to answer for the outputs of its chatbot was not decided by the court. Anyone unwilling to bear the risk should negotiate it into the contract: warranties regarding the training data and an indemnity in the event of infringements.

For companies that merely use AI content, the judgment provides relief on one point only: the provider is liable for generating the output. Anyone who then publishes a text is using a third party's work and, as before, needs permission to do so. A chatbot is not a licence, and that applies to song lyrics just as much as to anything else a model might have learned by heart.

It remains open whether the decision will stand on appeal before the Higher Regional Court of Munich. The same chamber did, however, decide in the same direction against the music generator Suno in July 2026.