On July 31, 2026, the Munich Regional Court I (Landgericht München I) largely upheld claims for injunctive relief, disclosure of information, and damages in a lawsuit brought by the music rights organization GEMA against Suno. The case number is 42 O 763/25, and the ruling has not yet become final. At the core of the court's reasoning is the finding that six songs remained reproducible within Suno's v3.5 and v4 models, and that creative elements of the original works appeared in the outputs. The German court treated memorization within the model as reproduction, rejecting the text and data mining (TDM) exception. Regarding training conducted in the United States, the court evaluated the four factors of fair use—including the fact that the original works appeared in the outputs—and declined to find fair use.
The ruling does not draw a blanket line stating that "training generative AI is always illegal." Rather, it distinguishes between cases where a model learns features common to works in general and cases where it memorizes specific works to the point that they can be extracted. What this poses for AI developers is the challenge of being able to explain, in a continuous chain, everything from how training data was obtained to how it is retained in the model and how reproducible the outputs are.
Six Works: From Training Data to "Reproduction Within the Model"
The works at issue in the lawsuit are "Atemlos durch die Nacht," "Rasputin," "Big in Japan," "Forever Young," "Daddy Cool," and the refrain of "Mambo No. 5." What was disputed was the musical composition, not the lyrics. It was undisputed between the parties that these six works were included in Suno's training data.
According to the court, Suno extracted the songs from YouTube via stream ripping, copying them by circumventing "Rolling Cipher," a mechanism designed to prevent downloading of audio and video. However, the official summary of the ruling does not state that this circumvention was the reason for rejecting the TDM exception. The question of how the material was obtained and the question of whether the works remained in the model after training must be read separately.
What proved decisive was that the six works could be extracted as outputs from v3.5 and v4 running on servers located in Germany. The court described the state in which the content of training data was incorporated into the post-training parameters and could later be output as "memorization." Comparing the original works with the generated results, the court determined that, given the complexity and length of the songs, this could not be coincidental.
Based on this finding, the court held that memorization within the model constitutes reproduction under Section 16 of the German Copyright Act (Urheberrechtsgesetz). The TDM exception under Section 44b permits, under certain conditions, reproduction for the purpose of analyzing lawfully accessed works and extracting information such as patterns, trends, and correlations. However, the court held that retaining the original work in a reproducible state within the model exceeds the scope of information extraction.
Do 338 Generation Experiments Represent Ordinary Use?
The inputs used to demonstrate reproducibility included the original lyrics, musical style, and title of each song. Melody and harmony were not specified, nor were rhythm and arrangement given as input. Materials from the March 2026 oral hearing indicate that 176 attempts were made for "Atemlos," 124 for "Big in Japan," while only 4 were made for "Mambo No. 5." The total across the six songs reached 338.
The court characterized these as "simple prompts that did not constrain the result." However, an input that provides a song title and original lyrics narrows the target far more than an input that does not identify a specific work. Moreover, there is a large gap in the number of attempts per song, ranging from 4 to 176. The official summary of the ruling does not reveal how consistently reproduction was achieved with fewer attempts.
Even so, the court placed significant weight on the fact that results containing creative elements of the original works were obtained without inputting melody or rhythm. In other words, what this lawsuit demonstrated is not that "any input produces a copy," but rather the fact that when information identifying a specific work is provided, the content retained within the model substantially determines the generated result. If the full text of the ruling is published, a technical point of verification will be which outputs were selected from among the 338 attempts and to what degree of similarity infringement was found.
Suno, Not the User, Bears Responsibility
Suno countered that because the generated results arose from GEMA's deliberately targeted inputs, the user's actions severed the causal chain. It also argued that the model's weights are not a container storing the training data but represent statistical patterns and generalized features. According to this explanation, the similarity in output was the result of the input narrowing the search space and bringing learned correlations to the surface.
The court did not accept this argument. Given that Suno selected the training data, operates the model, and bears responsibility for the architecture and memorization, the court held that it was Suno that substantially determined the content of the output. It found that outputs generated within Germany constituted unauthorized reproduction and communication to the public, and that the act of providing the generative service itself infringes the right of communication to the public under Section 15(2) of the Copyright Act.
This part of the ruling narrows the conditions under which generative AI providers can disclaim responsibility by explaining outcomes as "depending on the user's input." At minimum, when a model can reproduce a specific work and the input that triggers such reproduction does not require complex creative direction, the service operator can be held accountable, including for the model's design and the selection of training data. That said, what the first-instance ruling addressed was v3.5 and v4, and the six works and outputs presented in this lawsuit. The conclusion does not automatically extend to all current or future models, or to every generated song.
Fair Use Also Rejected for Training Conducted in the United States
Another notable feature of this case is that the Munich court also examined reproduction that occurred during training carried out in the United States. The court grounded its international jurisdiction on Sections 131(1) and (2) of the German Collecting Societies Act (Verwertungsgesellschaftengesetz), which allows a collecting society to bundle related claims against the same infringer. It then applied U.S. law to the conduct that occurred within the United States.
The fair use doctrine under Section 107 of the U.S. Copyright Act weighs four factors in combination: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect on the market. The court concluded that in Suno's case, all four factors weighed against the company. Because the original works appeared in the outputs, the court reasoned that the reproduction during training could not be protected as a transformative use.
The official summary explicitly distinguished this case from the 2025 rulings in Bartz v. Anthropic and Kadrey v. Meta. In Bartz, the training use was found to be fair use on the premise that no exact copies of the plaintiffs' works were generated from the training. At the same time, the act of assembling pirated copies into a permanent library was found not to be fair use. In Kadrey, Meta prevailed because the thirteen plaintiffs failed to present sufficient evidence of market dilution, and the ruling itself explicitly limited its scope, stating that it did not find Meta's training practices in general to be lawful.
Rather than establishing a general principle contrary to these two U.S. cases, the Munich ruling incorporated a different piece of evidence—that the original works appeared in the outputs—into its fair use analysis. In generative AI litigation, the era of asking only "was training conducted?" is ending; outcomes are increasingly determined by the legality of how copies were obtained, the extractability of content from the model, and whether outputs substitute for the original works in the market.
The Shift Toward Licensed Models Faces a Test
Suno has already begun moving in a different direction. In November 2025, it announced a partnership with Warner Music Group to build a new-generation model that would "surpass v5," using high-quality licensed audio. It has also indicated plans to make features that use participating artists' names, voices, and songs opt-in, along with a plan to provide compensation. However, the partnership announcement did not set a fixed release date for the new model.
Eight days before the ruling, GEMA also launched "PLAI by GEMA," a training data product containing approximately 178,000 audio tracks spanning more than 60 genres. The product bundles audio together with metadata, copyright information, and master rights, and has secured the music analysis company Klangio as its first user. However, PLAI is designed as a production support tool intended to generate outputs that do not compete with the source tracks used for training—it is not a comprehensive license aimed at general-purpose music generation models like Suno.
The ruling is not yet final, and details such as the amount of damages and the specifics of any rejected claims cannot be confirmed from the official summary. Nevertheless, the findings regarding v3.5 and v4 have made concrete the legal cost of continuing to distribute a trained model as-is. If an appeal is filed, the key factors going forward will be whether a higher court upholds the link drawn between memorization and fair use, and whether Suno can clearly separate its previously announced licensed model from its older models.