Generative AI & copyright: German court rules that storing works in AI models constitutes infringing reproduction

Germany has reaffirmed its pioneering role in European litigation concerning generative AI. On 31 July 2026, the Munich Regional Court ruled on a dispute between GEMA (the German equivalent of SACEM) and the US company Suno (developer of a generative music AI tool).

What was at issue? The unauthorised reproduction of six musical works obtained by circumventing protection measures on YouTube (‘stream ripping’), for the training and operation of its generative AI models.

‘Storing’ = physical reproduction. Drawing on the analytical framework established in the LAION case and subsequently in GEMA v OpenAI (data extraction/conversion, training and enrichment, use via prompts), the judges held that the works are stored in a stable manner within the model’s parameters, materialised on servers in Germany. This fixation is equated with compression (such as MP3). The indirect perceptibility of the work via a prompt is sufficient to constitute an infringement of the right of reproduction.

Drawing on the case law of the CJEU (Pelham, VG Bild-Kunst, Mio and Konektra), the court applied the criterion of the work being recognisable to the average listener, irrespective of the overall impression produced, and concluded that there had been an infringement of the rights of reproduction and exploitation.

Unlike in the Getty Images v Stability case, the court rejected the argument that the repetition of identical prompts had artificially caused the infringement: on the contrary, this demonstrates that the work is indeed stored and accessible, with the prompts used having no influence on the melody or harmony.

No exception for text and data mining, nor host provider status. Whilst applicable in principle to the analysis of training data, it does not cover the storage of works within the final model, which goes beyond the mere extraction of statistical correlations. Furthermore, unauthorised access to the works deprives Suno of the benefit of the exception.

Suno is deemed to be the direct author of the infringement, exercising decisive control over the training and architecture of the model; the user plays only a passive role in triggering the process. The exemption regime for hosting providers (Article 6 of the Digital Services Act) is therefore excluded, as Suno generates its own content.

FIRSH’s inputs

Model-based memorisation is becoming the cornerstone of AI litigation. For rights holders, the ruling provides significant leverage; for providers of generative AI solutions, it requires a fundamental rethink of how training risks are managed and how datasets are traced.

Our firm/team can assist you in mapping your IP/IT risks, auditing your AI systems and negotiating your licence agreements.

Claire Poirson & Léa Monel