AI Training & Copyright Part 3: Recent Case Law by the Regional Court of Munich (“GEMA vs. Suno”)

The use of copyright-protected works for AI training has become one of the hot topics driving copyright discussions across the globe. There is much legal uncertainty, and a lively ongoing debate: This includes fundamental questions on whether and how AI training involves copyright-relevant acts, whether (and under which conditions) concepts such as fair use (US) or the text-and-data mining exception (EU) apply, as well as questions of international jurisdiction and applicable law.

The debate is particularly dynamic in the EU, with legislative activities and various courts cases in play (see our blog here). The Regional Court of Munich had made a – rightsholder-friendly – early attempt to untie the gordian knot in the GEMA v. OpenAI case (see our corresponding blog here) by assuming that the existence of copyright-protected material both in training data and AI output is evidence of the existence of a copyright-relevant reproduction inside an AI model. 

More recently, this court has issued its second judgment in a similar vein in the case GEMA v. SUNO proceeding. Not unexpectedly, the judgment is evolutionary concerning the court’s controversial “copy in the model” – theory, which is re-applied and extended in the SUNO case. But it also contains certain new lines of argument, particularly concerning international jurisdiction and the interpretation of US law. 

The dispute (case no. 42 O 763/25) concerned six musical works – “Atemlos durch die Nacht”, “Rasputin”, “Big in Japan”, “Forever Young”, “Daddy Cool”, and the refrain of “Mambo No. 5” which had been used to train SUNO’s AI music generator. GEMA prevailed on essentially all of its German-law claims and on its US-law claims relating to training on US soil: the court granted an injunction covering training-phase reproduction in the US, “storage” in the model’s parameters in Germany, and public performance of outputs and derivative adaptations in Germany, ordered SUNO to render an account of the relevant infringements and revenues, established SUNO’s liability for damages (largely from 1 July 2023), awarded pre-litigation legal fees, and ordered publication of the operative part of the judgment in the Süddeutsche Zeitung at SUNO’s expense. 

Key Considerations

The judgment concerns numerous interesting – and controversial – findings, including, the court’s assumptions that

(i)      German courts have international jurisdiction concerning AI training that exclusively took place in the US and can award a cease-and-desist order purely relating to activities on US soil, regarding US copyrights;

(ii)      US law is applicable to AI training in the US, and that such training does not constitute “fair use” if copies of training data are stored inside the model weights;

(iii)     it is proven that an AI model contains reproductions of training data if certain works were used for training data and such data can be extracted into the output of the AI model based on user prompts; and

(iv)     the AI model provider itself is responsible for infringing output, if such output is generated based on simple user prompts.

While it is only a first instance judgment which will likely be appealed, its experimental approaches to novel questions deserve a deeper look.

International jurisdiction

The court ruled on three distinct copyright infringement claims: Alleged infringements through AI training in the US, “storage” in the AI model weights, and the generation of infringing output. 

Assuming international jurisdiction for AI output and the alleged “copies in the model” was not particularly problematic in this case: The output was created within Germany, and the AI model in question was also hosted on German servers.

This is different for AI training: It was undisputed that all training activities took place in the US. Since copyrights are restricted to the territory of a specific country, a “German” copyright cannot be violated outside Germany. The AI training claim therefore exclusively concerned actions on US soil, based on US copyrights with no relevant connection to the territory of Germany. The court assumed international jurisdiction anyway: It relied on a peculiar provision in the German Act on Collecting Societies Act which foresees that CMOs may “bundle” claims against the same infringer before one court, if this court has jurisdiction for one of these infringements. The court argued that – since it had jurisdiction regarding AI output and its “copy in the model theory – it could also rule on the US infringement claim.

This interpretation is experimental and invites controversy: The provision was originally introduced to address the problem of “travelling infringers” in Germany (e.g. concert organizers who travelled from city to city) and thus arguably intended to regulate only local, not international jurisdiction. While the court correctly points out that – in other contexts – German courts have assumed that local jurisdiction also establishes international jurisdiction, this concerned cases where there was a (strong) connection to Germany. There is, however, no indication that the provision interpreted by the court allows German courts to rule on completely separate infringements without any tangible connection to Germany. The court seemed to be aware and tried to argue that a connection to Germany does exist, because GEMA is seated in Germany, and the affected authors are “predominantly” German. This reasoning is hardly convincing: The plaintiff’s seat is generally irrelevant in determining international jurisdiction. This also applies to the author’s nationality, because the copyrights in question are (exclusively) rights governed by US law. Ultimately, the court itself does not seem to consider the author’s nationality decisive, because it would otherwise have differentiated between “German” and “non-German” authors.

Infringement Analysis

As part of its infringement analysis, the court reapplies its (equally controversial)  “copy-in-the model” theory. The court finds that, since (i) the Suno AI models were undisputedly (also) trained with the respective songs, and (ii) very similar output could be produced by using “simple” prompts, there must be reproductions of these works inside the model. The court also finds that the text and data mining exception (§ 44b UrhG; Art 4 DSM Directive) does not apply to “copies in the model”. This is because the reproductions found inside the model do not serve any further analysis and there was no lawful access to the works. The court notes that the circumvention of YouTube’s rolling cipher feature constitutes a circumvention of a technical measure.

This line of reasoning is very similar to the court’s judgment in the GEMA v. OpenAI proceeding, raising the same – and further – questions, in particular:

• “Copy in the model”-theory

  As in the Open AI, case, the court did not appoint its own, independent technical expert on the AI model in question, relying instead on academic literature and the parties’ own (adversarial) technical submissions. This disregards that different types of AI models are trained (and operate) in a different manner. The court’s reliance on academic literature and party-submitted – rather than court-appointed, independent – technical evidence might limit its persuasiveness for other courts. Notably, in the Getty Images decision, the UK High Court arrived at a different conclusion and denied the existence of “copies in the model” after an extensive hearing of technical experts. 

   

  The lack of a technical analysis also makes it difficult to answer the legal question whether the model contains a reproduction of training data in the legal sense. It is an open question under EU law to which extent a “reproduction” requires a (certain) fixation. While the court – as a lower instance court – was not required to refer the case to the CJEU, it may have missed an opportunity to increase legal certainty at the EU level.

   

  Moreover, the court assumed that the AI output was based on “simple” user prompts, although the prompts used by GEMA were significantly more complex than in the OpenAI case: GEMA used the entire song lyrics for the prompts, and even provided additional input regarding the desired style (such as “80s, Synth Pop, male voices”) for the song “Forever Young”. Even then it took between 4 (“Mambo No. 5”) and 176 (“Atemlos”) attempts with identical prompts to arrive at the allegedly infringing output. It seems doubtful that this approach still qualifies as a “simple” prompt, rather than a targeted attempt to trigger infringing output. While GEMA claimed that the other attempts produced infringing output as well, it refused to put these attempts on file. The mere fact that the same prompt apparently led to up to 176 different results (“Atemlos”) casts doubt on the courts assumption that this is proof of a “copy in the model”. While the court seems to assume that the neural networks in AI models are “deterministic” and that a variation of the output is artificially created at a later stage of data processing, this finding is based on little technical analysis. The somewhat artificial distinction between a “deterministic” model and an element of chance which is added at a later stage seems underexplained as well.

• Responsibility for AI Output

  The same open questions also affect the analysis regarding the responsibility for AI output: the court finds that SUNO is responsible for infringing output, because it is the model provider, has used the works at-issue for training purposes and controlled the training. It assumes that the prompts by the users were “simple” and “open-ended” which means that liability did not shift towards the user. This general allocation of liability is in line with the GEMA v. OpenAI judgment where the court also assumed that OpenAI is not merely providing technical services. 

   

  However, the considerable complexity of the prompts used in this case, the number of attempts, as well as the apparent intent to create copyright-infringing content stretch the limits of the court’s assumption quite far. These factors speak more in favour of a process which was steered by the user, not SUNO – and consequently direct liability of the user rather than the provider.

• Lack of external expertise

  The court’s reluctance to draw on external expertise was also not only confined to the central point of a technical analysis of the AI model. Equally, it did not appoint independent experts in two further areas which proved to be central for the decision, i.e. musicological, and US legal expertise.

   

  No musicological expert

  The court did not appoint an independent musicological expert, reasoning that the judges themselves belong to the musically-informed public, although it did consider competing private expert opinions submitted by both parties. The corresponding statements in the judgment are quite extensive, and require deep understanding of musical theory and common tropes for specific music genres (e.g. “The characteristic progression, the dramatic gesture, and the harmonic resolution remain recognizable, even though the arpeggio figure (referred to by the defendant as an ‘arpeggiator figure’) is in F-sharp major rather than E major, as in the original. The striking choice to end the chorus phrase in G-sharp minor—even though a major resolution would be the natural choice in Schlager music—is also retained."). The judgment does not explain that the court had the necessary own expertise to arrive at such detailed conclusions.

   

  No US law expert

  Crucially, the court also did not appoint its own, independent expert on US law, in particular concerning the notoriously difficult and nuanced fair use analysis, but instead applied its own interpretation of US law. Unlike the controversial assumption of international jurisdiction, the application of US law to training occurring in the US is a consequent application of the principle of territoriality. Unfortunately, however, the lack of external expertise may well have impaired the quality of the court’s analysis:

  While existing US caselaw (in Bartz and Kadrey) found AI training to be fair use on the facts before those courts (both decisions turned on the absence of substantially similar infringing output) the court disregards these precedents arguing that these did not concern cases of “copies in the model”.

  The US fair use test looks at four factors: 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 potential market”. The court concluded that none of these factors operated in favour of a fair use here. It added that the novelty of the AI technology did not compel a different result, since not all models memorize their training data and rejecting fair use here would not impede progress but merely require licensing of protected training data.

  While a full assessment of the court’s fair use analysis goes beyond the scope of this blog, the court’s assessment rests on a thin evidentiary basis, in particular regarding factors one and four: The court’s assumption that the use was not transformative may be overly focussed on the individual work, rather than the overall – clearly highly transformative – use case of the AI model. Moreover, the court’s assumption regarding the effects on the potential market seems to lack any robust and representative evidence, falling short of the standards for a state-of-the-art assessment of this factor by a US court.

What’s next?

After GEMA v. OpenAI – the Munich court’s general stance is not entirely surprising. However, the SUNO decision pushes its position even further. The decision clearly shows that the court’s direction was not a “one-off”. Rightsholder may therefore increasingly try to bring cases before the Regional Court of Munich. At least two further cases (concerning children’s books series “Der kleine Drache Kokosnuss” and “Das NEINhorn”) are already pending. 

However, reports about the “groundbreaking” character of the decision should be taken with a grain of salt:

The court’s position is novel and experimental in many respects. As such, it is an interesting contribution to an ongoing debate. The questions decided by the court will very likely be dealt with by the Higher Regional Court of Munich, the German Federal Court of Justice, and the CJEU. 

Regardless of whether the court’s view ultimately withstands further review, its approach has significant limitations:

• The court’s assumption of international jurisdiction for US training is based on a bold interpretation of a provision specifically for collective management organisations. Other rightsholders cannot rely on this provision.
• The court’s theory on “copies in the model” requires the demonstration of memorization/ regurgitation on the level of individual works. The Suno judgment (137 pages for just six songs) shows the limitations of this approach: It will be difficult to produce the necessary evidence for larger corpuses of work instead of just a few test cases. If AI providers implement more reliable output-filters in the future, demonstrating the existence of “copies in the model” may also become even more difficult.

Bottom line

While the Regional Court’s judgment is remarkable, it is ultimately (only) another mosaic piece in a quickly developing EU litigation landscape. The discussion – which will also be heavily influenced by political and regulatory developments – is far from over. One thing is certain: These are interesting times for copyright enthusiasts.