The Munich Regional Court I held on July 31, 2026, that training an AI model with copyrighted works without license infringes copyright if the AI model reproduces these works in outputs on relatively simple prompts. The court found that the model memorizes the works within the model as reproduction, and that such reproduction is not covered under the TDM exemption.
The legal dispute between Germany’s collecting society GEMA and Suno, the U.S.‑based provider of an AI music generator, concerned six well-known musical works – including “Atemlos durch die Nacht,” “Daddy Cool,” “Rasputin,” “Big in Japan,” and “Forever Young,” which Suno had undisputedly included in the training data for its v3.5 and v4 models. The recordings were extracted by stream ripping from YouTube, circumventing the rolling cipher, which is a technical protection measure. GEMA generated outputs in which protected musical elements of the originals were recognizable. The court granted injunctive relief and held Suno liable in damages in principle, the amount still to be determined.
The Chamber was convinced that the protected musical works were reproducibly contained in the models stored on German servers. Reproduction relevant under copyright law does not require that a work be stored as a distinct audio or other file. It is sufficient that the work is fixed in the model parameters in a form that can be made perceptible by technical means.
The court inferred the memorization from the substantial similarity between the training works and the generated outputs. Given the complexity and length of the matches, it ruled out the possibility of random generation. Suno argued that its model weights contain only mathematically learned patterns rather than copies of specific works, and that the similarity between outputs and originals was caused by GEMA’s deliberately narrowed prompts rather than by memorization. The court rejected both arguments: Suno selected the training data, designed the model architecture, and caused the memorization, while the prompts gave no concrete musical instructions.
GEMA entered identical prompts between four and 176 times, depending on the work. Each contained the lyrics, the title, and a stylistic instruction, but no specification of melody, harmony, rhythm, or arrangement. The court therefore treated the prompts as simple and open-ended; repetition alone did not steer the outputs. This finding is likely to be a focal point on appeal.
The court held that the TDM exception under Sec. 44b of the German Copyright Act (UrhG) did not apply because memorization served no further analytical purpose but resulted in an additional, permanent reproduction. The court further held that circumventing the rolling cipher meant that Suno lacked lawful access to the works. The court’s reasoning rests on several distinct but reinforcing grounds.
The core of the TDM exception – both under Sec. 44b UrhG and its EU law basis in Article 4 DSM-Directive – is that reproductions must be made “for the purpose of” text and data mining. Reproductions within the model (Phase 2) do not serve any further data analysis purpose; they are an end result, not a preparatory step. The court therefore held that the purpose requirement is simply not met.
The court drew a clear structural distinction between two phases:
The TDM exception was designed for Phase 1 only.
TDM, by definition, targets the extraction of “information” – patterns, trends, correlations – from works. The works themselves are not supposed to be reproduced. Where memorization occurs, however, the work is not merely analyzed; its content is taken up into the model parameters in a way that allows it to be reproduced in outputs. This exceeds what the exception permits.
The court also rejected an analogical application of Article 4 DSM-Directive to Phase 2 reproductions. Even if one assumed a gap in the legislation, the comparable interest situation required for analogy is absent: the TDM exception rests on the premise that exploitation interests are not affected, because only bare information is extracted. With memorization, exploitation interests are directly and substantially harmed. Extending the exception by analogy would leave rights holders unprotected without any compensation – contrary to Recital 17 DSM-Directive, which requires that harm to rights holders from exceptions be only “minimal.”
Separately and independently, the exception was also unavailable because Suno had circumvented the rolling cipher technical protection measure to obtain the works from YouTube. Article 4(1) DSM-Directive and Sec. 44b(2) UrhG both require “lawful access” to the work as a precondition. Circumventing a technical protection measure to download works that could only be streamed freely does not constitute lawful access.
Suno argued that denying the exception would unconstitutionally restrict a legitimate business model protected under freedom to conduct a business (Article 16 EU Charter; Article 12 GG). The court rejected this; a business model premised on using others’ intellectual property for free is not recognized by either legal order. The obligation to license is, conversely, required by Article 17(2) EU Charter, which explicitly protects intellectual property.
The court held that compliance with the copyright-policy and training-content-summary obligations under Article 53(1)(c) and (d) AI Act does not replace permission as the AI Act creates no new copyright exception.
Article 53(1)(c) AI Act obliges providers of general-purpose AI models to put in place a policy for complying with Union copyright law and, in particular, for respecting opt‑out reservations under Article 4(3) DSM-RL. The court held that this provision refers to the DSM-RL and does not define or expand its scope; it does not extend the TDM exception to cover reproductions within the model (Phase 2). Recital 108 AI Act confirms this: the AI Act does not affect the enforcement of Union copyright rules.
Article 53(1)(d) AI Act requires providers to draw up and publish a sufficiently detailed summary of the content used to train the model. The court rejected Suno’s reading that publishing such a summary exhausts all copyright obligations. The purpose of the summary, as stated in Commission Communication C(2025) 8311 final, is to help rights holders obtain relevant information and to facilitate the exercise and enforcement of their rights – not to replace the need for a license. Listing data obtained by crawling and scraping in the summary does not render that use lawful.
The court further noted that the Code of Practice on general-purpose AI models explicitly states that adherence to the Code “does not constitute compliance with Union law on copyright and related rights.” Accordingly, neither the AI Act obligations nor the Code of Practice constitute a copyright exception, and unlicensed use of protected works in AI training remains subject to the ordinary rules of copyright law.
The outputs also infringed the rights in the musical works because they recognizably incorporated protected musical elements of the originals. The court attributed these infringements to Suno rather than its users: Suno had selected the training data, trained the models, and caused the works to be memorized. As the outputs constituted Suno’s own information, the court held that the hosting safe harbor under Article 6 DSA did not apply. Particularly far-reaching is its finding that the mere offering of the model and the music-generation application infringed an unenumerated right of communication to the public under Sec. 15(2) UrhG.
The Munich Regional Court examined the four factors under 17 U.S.C. Sec. 107 and rejected Suno’s fair use defense. Central to the court’s assessment was that musical works with substantial similarities to the preexisting works could be generated using the AI model – a difference to other notable U.S. cases, in which this had not been the case. The court also considered the unlawful provenance of the training copies.
The judgment is not final. Pending lawsuits such as the CJEU Case C-250/25 (Like Company) may provide further guidance on memorization, reproduction, and the scope of the TDM exception.The Advocate General’s opinion is expected on September 3, 2026.