Recently, a court in Munich made an important ruling against the well-known AI music generator Suno. The court determined that Suno committed copyright infringement both in the training process of its AI model and in the final output, and formally rejected the company's defense of fair use. This decision is not yet final.

This lawsuit was initiated by the German music copyright organization GEMA, and the case centered around six well-known songs, including Kristina Bach's "Atemlos durch die Nacht" and the classic song "Rasputin." After reviewing the case, the court pointed out that versions 3.5 and 4 of Suno's model could accurately reproduce original elements from these famous songs. Technically, this phenomenon is called "memorization," where the model not only absorbs general patterns during training but also stores specific original content, which can be retrieved later as generated results.

Regarding Suno's argument that it only stored mathematical patterns and that the outputs were attributed to user prompts, the court did not accept this claim. The court clearly stated that since Suno selected the training data on its own and decided the model's architecture, it should bear direct responsibility for the infringing outputs, rather than shifting the blame to users who provided open-ended prompts. Additionally, the exception clause related to text and data mining in Germany does not apply to this "memorized copying" behavior.

In addition, the court exercised jurisdiction over training activities that occurred within the United States and, based on U.S. law, determined that the fair use doctrine does not protect Suno. The court noted that unlike some U.S. cases previously deemed transformative, Suno's model can directly generate results highly similar to original works with simple input.

Aside from the issue of similarity in the output, the court also noticed a key detail during the trial: Suno used streaming scraping technology to extract music from YouTube and bypassed technical protection measures designed to prevent downloads. Currently, this ruling is still under further review, and it may have a profound impact on copyright compliance and the sources of training data across the entire AI industry.