On 31 July 2026, the Regional Court of Munich I ruled in GEMA v. Suno. On 2 August 2026, the transparency obligations under Article 50 of the EU AI Act took effect. Both address the same underlying question from opposite directions: what must be known, and provable, about how a piece of audio came into existence?
This piece sets out the factual position — without passing judgment on the technology, but with attention to the operational requirements now emerging for rightsholders, producers and providers.
01The ruling: what was actually decided
The 42nd Civil Chamber of the Regional Court of Munich I largely upheld GEMA's claims for injunctive relief, disclosure and damages (case no. 42 O 763/25). Six works were at issue, including Atemlos durch die Nacht, Rasputin, Big in Japan, Forever Young, the chorus of Mambo No. 5 and Daddy Cool. As far as can be established, this is the first court decision worldwide concerning AI-generated audio content.
The core reasoning distinguishes two separate acts:
- the storage of the works within the model — treated by the court as an interference with the reproduction right
- the output to users — treated as an interference with the right of communication to the public
In doing so, the chamber continued the line it had taken in its OpenAI ruling of November 2025, where it had already interpreted the text and data mining exception narrowly.
The evidence was decisive. GEMA was able to document that the system produces output largely corresponding to the original works in melody, harmony and rhythm. The dispute therefore did not turn on the abstract permissibility of machine learning, but on demonstrable output behaviour.
The international dimension is the less-discussed and possibly more consequential part: for the first time, a European court ruled on a training process that took place in the United States — concluding that a licence would have been required under US copyright law as well.
The judgment is not final. Suno rejected it. The court's press release does not specify on which points GEMA did not prevail. Parallel proceedings brought by Universal Music and Sony Music are pending in the United States.
One detail is notable for the procedural history: Suno had already conceded that it trained on GEMA repertoire, disputing only the obligation to pay remuneration.
02The regulation: what Article 50 requires
The EU AI Act's transparency obligations have applied since 2 August 2026. Public discussion frequently reduced this to the claim that all AI-generated content must now be labelled. That is not accurate.
The regulation addresses defined risk situations — primarily deception about origin or authenticity. Two levels matter for audio:
Providers of generative systems must ensure that generated audio, image, video and text content is marked as synthetic in a machine-readable format. The technical implementation — watermarking, metadata, or other methods — is not prescribed in detail; what is required is clear and unambiguous information.
Deployers publishing AI-generated content must make this perceptible to users where a deception scenario arises. For text, an exemption applies where qualified human editorial review has taken place and responsibility has been assumed — cursory sign-off does not suffice.
User-facing disclosure. For providers already on the market, a grace period for machine-readable marking runs until 2 December 2026.
Fines of up to EUR 15 million or 3 per cent of worldwide annual turnover, whichever is higher.
03The gap between the two
The ruling and the regulation intervene at different points. The ruling concerns the input side: what material was permitted to enter a model? Article 50 concerns the output side: is it apparent that content is synthetic?
Neither governs what lies in between — documentation within the production chain.
A practical illustration: an agency produces a commercial. Part of the music comes from a generative system, part from a composer. One voice was recorded, a second was synthesised. The spot is delivered.
Under Article 50, the synthetic portion must be identifiable. Following the logic of the Suno ruling, the system used must have been trained on a licensed basis. Both are difficult to establish retrospectively if nothing was recorded during production. The questions that arise in any review situation — which system was used, in which version, on what data basis, which parts of the output were subsequently reworked by a human — are hard to reconstruct without contemporaneous documentation.
This is not a legal subtlety but an operational problem. It affects not the model providers, but the companies incorporating their output into their own products.
04Unresolved: material introduced by users
A second area remains open. Many platforms allow users to upload their own audio — as reference, as a voice sample, as a starting point. Terms of service routinely assign responsibility to the user: by uploading, the user warrants that they hold the necessary rights.
This construction is standard in contract terms. In practice it shifts the duty to verify onto the party with the least information and the least capacity to check. A private user typically knows neither the rights chain of a recording nor how their upload will be used downstream.
The Suno proceedings did not concern this scenario — what was at issue there was training carried out by the company itself. No comparable decision yet exists on how protected material enters systems via user uploads, or who bears responsibility when it does.
05Economic assessment
Commercial momentum continues regardless of the legal position, and it is explicable.
Generative systems produce continuously, in any number of languages and for any number of markets, without the constraints of human production. For catalogue segments with high demand and low differentiation requirements — background music, advertising production, functional music — this represents a substantial cost advantage.
There is a further observation, uncomfortable for the creative sector but empirically relevant: a significant share of the audience does not ask how content was made. Preference forms around the result.
Rightsholders are therefore responding on two tracks. Alongside litigation sit cooperation models: Udio has agreed with Universal Music on a joint platform. The Suno case can be read in the same frame — the proceedings followed a licensing offer from GEMA that was not accepted.
A market differentiation between providers with a licensed data basis and those without. For buyers — labels, agencies, broadcasters — the licensing position of a system becomes a selection criterion alongside quality and price.
06Opportunities
The development opens up several positions that did not exist two years ago.
For rightsholders, an enforceable claim to remuneration for the use of repertoire in training now exists. What was previously regarded as practically unenforceable is the subject of a ruling — with a disclosure claim that provides the basis for quantification.
For providers with a clean data basis, compliance becomes a commercial argument. A system whose training basis is licensed and documented holds an access advantage in regulated markets that retrospective clean-up can hardly replicate.
For European market participants, the competitive logic shifts. In the race for model scale, positions are largely allocated. In the field of demonstrably compliant, specialised systems they are not — and the regulatory requirement treated elsewhere as a constraint is already part of the home market here.
For producers and studios, a differentiator emerges. Being able to evidence how a production came about serves a demand that is likely to grow rather than shrink as synthetic content spreads.
07Open questions
The unresolved points are equally clear:
- Finality. The judgment is a first-instance decision. The appeal is pending, as are the US proceedings.
- Technical standards. Article 50 requires machine-readable marking without prescribing a method. How robust existing watermarking approaches are against conversion, compression and further processing remains open.
- Hybrid productions. Where material moves repeatedly between human and machine processing, it is unclear at what threshold a disclosure obligation arises.
- User uploads. The allocation of liability has not been judicially clarified.
- Enforcement. How supervisory authorities will operate in practice, and what evidence they will require, will only become apparent in application.
08Assessment
The two developments of this week do not mark a rupture but a clarification. Generative audio systems are not being called into question — the requirements regarding their basis and their labelling are becoming more concrete.
For practice, this implies a shift: provenance has until now been a matter of trust between client and producer. It is becoming a matter of evidence — towards rightsholders, supervisory authorities and buyers.
What is missing are workable methods of producing that evidence without burdening the production process. That is what we work on at URSPUR: criteria and forms of documentation that make the origin of audio content demonstrable — technically, verifiably, and capturable in ongoing operation.
We will continue to monitor developments, in particular the appeal proceedings, the implementation of machine-readable marking by December 2026, and the question of user uploads.