GEMA v. SunoAI: the second German ruling on copyright and generative AI training

Two sets of issues: the input side and the output side

The ruling addresses several issues. First, however, it is important to distinguish between two different levels of analysis:

1. The input side: what is used to train the model.
2. The output side: what the model produces in response to the user’s prompt.

The first issue concerns memorisation. SunoAI not only used the songs as training material, but also memorised their content – sounds, lyrics and melodies – retaining them in the model in their entirety and without modification.

Memorisation frequently causes another phenomenon: the so-called “regurgitation”. Given an appropriate prompt, the model “regurgitates” parts of the original song, or even the entire song, precisely because it has been memorised in the training dataset.

According to the German Court, determining whether memorisation has occurred is straightforward: the output can simply be compared with the protected work. If they match, even partially, it can be inferred that the song was used as training material.

In this case, the fact that the model produced outputs that were identical, even only in part, to the songs at issue shows that those songs were improperly used as training material, without any prior authorisation from their copyright holders.

It was also demonstrated that the original song was not “regurgitated” randomly, as an accident or hallucination. It was consistently produced as output on several occasions.

Under the German ruling, this constitutes an infringement of the copyright in the six songs.

Territoriality: the German Court also applies US law

The Court therefore ruled on copyright infringement under both the German Copyright Act (Urheberrechtsgesetz) and the U.S. Copyright Act, 17 U.S.C..

One particularly interesting aspect concerns territoriality: the German Court held that it had jurisdiction not only over copyright infringements occurring in Germany, but also over infringements committed in the United States during training.

Consequently, the Court was able to apply both German law and the doctrine of “fair use”, which is used in the United States to assess whether a use is exempt from copyright infringement liability.

This is crucial, since the model was trained in the United States. It is closely connected to the two sets of issues identified above:

1. On the input side, the Court assesses whether using the songs as training material in the United States was lawful.
2. On the output side, the Court must consider whether the generative AI’s reproduction of a song in response to a user’s prompt, which occurred in Germany, infringes the reproduction right.

Does the EU Text and Data Mining exception apply?

Central to the ruling are the considerations concerning the applicability of the Text and Data Mining exception under Directive (EU) 2019/790, known as the “DSM Directive”, particularly Article 4. On this point, see our previous article. The Court rules out its applicability for several reasons:

1. The songs are memorised in the dataset, rather than being used only temporarily for the sole purpose of analysing information or data.
2. Access to the songs is not lawful, since they are used without GEMA’s prior consent or an agreement with the organisation, and are downloaded by circumventing security measures.
3. GEMA and YouTube had expressed their reservation of rights (opt-out) in a readable form, as required by the DSM Directive.

How is the US fair use doctrine applied?

As regards the training that took place in the United States, the German Court applies US law as the lex loci protectionis—the law of the “country for which protection is sought”—as required by Article 8(1) of Regulation (EC) No 864/2007, known as the “Rome II Regulation”. The Court considered itself able to decide the case under US copyright law without the assistance of an expert, which the defendant had requested.

The Court had to apply the four factors of the US fair use test (see Section 107 of the U.S. Copyright Act, 17 U.S.C.), which is used to assess whether the use of a copyright-protected work constitutes “fair use” or copyright infringement. Examining each factor in turn:

1. The purpose and character of the use are not transformative, as required by established US Supreme Court case law, because the work is used solely to “regurgitate” an output that is substantially identical to the original song.
2. As regards the nature of the copyrighted work, copyright protection is fully applicable in this case, since the songs are protected works in the United States.
3. The amount and substantiality of the portion used weigh against fair use, since the entire song was used.
4. The effect of the use on the potential market for the protected work weighs against the defendant, because the songs produced as output target substantially the same market as the input.

What did the Court decide?

On the output side, the reproduction right was held to have been infringed, because the reproductions were neither incidental nor transient. They also had economic significance, meaning that the exception under German law did not apply.

The prompts used were simple and were not specially engineered to elicit the songs as output. This therefore provides clear evidence of memorisation: the song could not have been “regurgitated” as output unless it had been used as training material.

For these reasons, the Court granted GEMA an interim injunction against the defendant, prohibiting it from:

1. Reproducing the songs without GEMA’s consent for the purpose of training generative AI models in the United States.
2. Reproducing them within the model itself in Germany.
3. Publicly performing the songs through Suno’s platform.
4. Committing any infringement connected with distribution.

The Court also ordered Suno to disclose to GEMA the extent of the unlawful acts of reproduction of the musical works and to pay damages, which have yet to be quantified.

What are the practical implications of the ruling?

For the second time, the same German Court has declared the use of copyright-protected songs to train generative AI unlawful. The substantially identical outputs were also found to constitute copyright infringement.

From a global perspective, this establishes a clear principle: according to the Munich Regional Court I, copyright-protected material cannot be used indiscriminately to train AI models. No exception, including the Text and Data Mining exception, can justify memorising and indiscriminately using protected works.

This would leave companies providing generative AI models with only one available path: licensing agreements with copyright holders.

This point strongly favours creators, who have welcomed the approach taken by the German Court. For AI system providers, however, it presents greater difficulties, given the enormous quantities of material they use indiscriminately.

For now, several possibilities remain open. Courts around the world have adopted different approaches.

There is still no clear, uniform position. We must wait, particularly in the European Union, to see how other courts will rule on the issue—including in Italy, where the first case on this matter involves Perplexity AI Inc., sued by RTI and Medusa Film, both part of the Mediaset Group.

Nevertheless, the two GEMA rulings remain an important initial milestone for creators—and a major problem for AI companies.