ChatGPT infringes copyright of German songwriters, according to the Munich regional court

The court’s reasoning
In the case, GEMA argued that OpenAI had unlawfully used the works of GEMA’s clients to train ChatGPT, and that the AI model produces outputs which are highly similar to the original works of the authors. OpenAI, on the other hand, argued that the text and data mining (TDM) exceptions justified any use of the original works to train ChatGPT. The TDM exceptions provides for a right of reproduction of works for the purposes of analysing text and data to obtain information. Furthermore, OpenAI argued that to the extent any outputs resemble original works, this is not the responsibility of OpenAI but instead the user whose prompts generate the output.

The court found that OpenAI had used the works to train ChatGPT, and that ChatGPT generated outputs which highly resembled the original works. In discussing whether the text and data mining exceptions were applicable, the court noted that the works had been «memorized» in the AI model, which enables the AI model to generate similar outputs long after the training has been concluded. The court noted that the text and data mining exceptions do permit the reproduction of works for the purposes of training AI, but that this exception does not permit a «permanent» reproduction of the work, resulting in the AI model being able to reproduce new copies of the work indefinitely.

Against this background, the court found that OpenAI’s use of the works were not permitted by the text and data mining exceptions. Furthermore, the court rejected OpenAI’s argument that the liability for the use remains with the users who prompt the works.  Accordingly, the court held OpenAI liable for infringing the German songwriters’ copyright and ordered OpenAI to cease activities relating to the authors and pay damages.

The decision is not yet publicized, and this is based on the court’s press release (available here). The decision is not final and may be appealed.

Implications for the TDM exceptions
The text and data mining exceptions (TDM) were introduced in the Copyright in the Digital Single Market Directive (Directive (EU) 2019/790), which is not yet incorporated in Norway. One of the legislative considerations for introducing this exception has been namely to facilitate the training of AI models (cf. for example the Norwegian consultation paper on the implementation of the DSM Directive). However, if the result from the Munich regional court is upheld, it seems unclear whether the TDM exception is sufficient for all aspects of training and offering of AI models. The final decision in this case will likely influence other courts in other EEA/EU member states in interpreting and applying the TDM exceptions.