AI has increasingly been used to generate music and entertainment content, but how are applications like Suno and Udio able to generate ‘new’ works? As former MP, Kevin Brennan, explained, AI learns ‘the patters of human creativity ... based on algorithms and predictions’, in order to generate new content. The problem this poses for the entertainment and music industries is that without licenses, AI companies would be exploiting and damaging the industries’ very foundation: Intellectual Property (IP).
Copyright law specifically, serves as an integral component of a functioning creative landscape. From songwriters to film and tv producers to authors and more, rightsholders hold the power to exploit their IP and capitalise on it. When a legal person uses copyrighted work without the authorisation of all the rightsholders, there may be copyright infringement. In the context of AI training on copyrighted works, copyright infringement is of major concern globally.
There is significant legal tension between AI companies and record labels, production companies, artists and other rightsholders in the entertainment and music industries. There have been a number of lawsuits and licensing deals recently, however the Munich Regional Court’s ruling at the end of July 2026, marks an important landmark for the problem of AI training on unlicensed works.
The Facts of GEMA v Suno Inc:
GEMA, the German music collecting society, brought an action against Suno, who provides an AI music generator, in January 2025, after Suno failed to respond to GEMA’s request for licensing. The case concerned six compositions which GEMA had in its repertoire: “Daddy Cool”, “Mambo No. 5”, “Forever Young”, “Atemlos dur die Nacht”, “Big in Japan” and “Rasputin”. In bringing the lawsuit, GEMA sought an injunction against four copyright rights-related uses of the works.
In the US, Suno trained its AI models on copyrighted works, including those in dispute. The works were obtained by stream-ripping on YouTube, which involves going around YouTube’s downloading restriction. This means that Suno did not acquire licensed music to train its AI, and this was not in dispute in the case. After obtaining unlicensed musical works, Suno attached metadata to those audio files and converted them to numerical units which Suno’s AI model’s parameters trained on. However, the final model, which is stored in Germany, does not itself contain the original audio files.
The issue in this case was not establishing if Suno had in fact trained on copyrighted works without a license, rather, it was that Suno argued that it did not need a license in the first place. Suno challenged the jurisdiction of the German court because the training was done in the US. Additionally, Suno tried to rely on both US fair use defence and the EU’s text and data mining exception. The ruling in this case is therefore of high importance in the current increasingly technological landscape.
The Munich District Court’s Ruling:
Firstly, it should be noted that the German court was able to rule on the training of the AI model even though it happened in the US, because of provisions in the German Collecting Societies Act, s 131(1)-(2), and the German Code of Civil Procedure, s 32. There are restrictions to this as only collecting societies may benefit from this law and there must be an infringement that took place in Germany.
The court found that there was reproduction of works because the AI model memorised them and exact copies were not needed. By inputting a style with lyrics and a title and comparing the AI outputs to the original works that were the subject of the case, memorisation was proven. Notably, Suno argued that the AI-generated infringing outputs were the user, GEMA’s, responsibility because the prompts that were given to the AI were “complex”. The court rejected this. The prompts were not close-ended and the AI model was free to decide important aspects of the final output: the harmony, tempo, melody and rhythm. Using the reasoning of the GEMA v OpenAI case that the same court ruled on previously, the works were fixed in numerical probabilities and exemplified in parameters that were specified, and so they were reproduced.
The court determined that there was no text and data mining exception in this case. The reproductions were not for data analysis and Suno, by going around YouTube’s downloading restriction, had obtained the works unlawfully. Thus, no text and data mining exception could be relied on by Suno.
The German court also ruled on whether the training of Suno’s AI model in the US infringed US copyright law. In doing so, reproduction was again established and the court analysed if the four fair use factors (the work’s nature, the purpose but also character of the use, the amount of the works used, and the market impact (17 U.S.C. § 107)) could be weighed against Suno. Importantly, the court distinguished the facts of GEMA v Suno, with Bartz v Anthropic, because in Bartz the AI did not output its training data. The court found that none of the four factors aided Suno in proving that they did not need a license. Their use of the works was commercial not transformative and there was bad faith in acquiring the music unlawfully. The compositions were fully copied and Suno had taken the expression of the works.
Why is this Important? Should the Courts of other Nations also hold that Training AI Requires Licenses?
GEMA v Suno is a landmark victory for creatives and other Nations should consider requiring AI companies to obtain licenses to train their models on copyrighted content. The ruling on US fair use does raise the question: what would a US court have thought? If some record labels like Warner Music Group did not settle in their infringement suits against Udio (another AI music generator) and Suno, with licensing deals, a US court may have applied the same analysis as the German court in GEMA v Suno. Fortunately, there is still a RIAA (Recording Industry Association of America) coordinated lawsuit against Sony Music and Universal Music Group and other litigations in the US. This means US courts can adopt the German court's reasoning, and if they do, those across the music, entertainment and publishing sectors that have licensed may have chosen to do so too soon.
Additionally, it is hopeful that European nations also find infringement and that a text and data mining exception is not available to AI companies that unlawfully obtain copyrighted material to train their AI models. While there is hope for creatives to protect their work from unlicensed AI training, the factual nuances of GEMA v Suno, namely that the case focused on only six compositions and there was proof that Suno could reproduce those exact compositions, should be noted. Where future cases have a significantly larger volume of copyrighted works in question, it will be harder to evidence reproductions of all works in question. Nevertheless, the German case provides optimism for the protection of creative works, which is crucial for the entertainment and music industries as IP is their foundation.
Sejal Patel, August 2026
Sources:
APPG on Music Report: ‘Artificial Intelligence and the Music Industry – Master or Servant?’ https://www.ukmusic.org/wp-content/uploads/2024/04/APPG-AI-Report-Low-res.pdf
https://www.musicbusinessworldwide.com/suno-infringed-copyright-in-gema-case-german-court-rules/