German court rules out free use of music for AI: what does the GEMA judgment mean for authors?

Six compositions from GEMA's repertoire surfaced in music that Suno generated, although the prompts had not asked for them. For the court, that pointed to a model which had retained parts of those works. When does training become an act of reproduction under copyright law, and what does that mean for authors who see their work disappear into datasets?

May an AI company use existing music to train a model without asking permission or paying the people who made it? That question has been central to the debate about generative AI and copyright for some time. A recent judgment of the Regional Court of Munich I (Landgericht München I) in the dispute between the German collecting society GEMA and the AI music service Suno gives an important answer.

On 31 July 2026 the court ruled largely in GEMA’s favour. In its view, Suno infringed copyright in six well-known compositions from the GEMA repertoire. Among them are Forever Young by Alphaville, Mambo No. 5 by Lou Bega and Daddy Cool by Boney M.

For authors and other rightholders, what lies behind that finding matters most. The judgment concerns more than six songs that happened to resemble AI output too closely. The court addresses a more fundamental question: what happens, as a matter of copyright, when protected work is used to build a generative AI model?

From training data to new music

Suno is a generative AI service that lets users create music from textual instructions. Behind that relatively simple user experience sits a model trained on large quantities of existing music.

It was established in the proceedings that Suno had trained its models in the United States on millions of complete sound recordings, and that the six compositions from the GEMA repertoire were among them. According to GEMA, no licences had been obtained.

The presence of protected material in the training data was not the only relevant point, however.

GEMA played the court a number of tracks Suno had generated. Melodies, harmonies and rhythms recurred in them which, according to GEMA, closely resembled the original compositions. What stood out was that generating those tracks required no elaborate musical instructions. The prompts contained a lyric, a style indication and a title, but did not prescribe the melody, the harmony or the rhythm.

That mattered to the court. Where such musical elements nonetheless resurface in the generated music, it may indicate that the model has not merely learned general patterns from the training data, but has retained parts of specific works.

When ’learning’ becomes reproduction

This brings the judgment to a question that plays out far beyond the music industry.

AI providers often stress that a model does not simply store the training data as a collection of files. During training, data is converted into model parameters, numerical values with which the system can then generate new output.

The court accepts that technical distinction, but does not draw from it the conclusion that copyright falls out of the picture.

According to the court, there can be a reproduction in the copyright sense where a protected work has been ‘memorised’ in a model’s parameters. That does not necessarily require a classic copy of the music file to be present in the model. What is decisive, in the court’s view, is which protected elements of a work have been taken over into the model.

For rightholders that is a substantial point. The legal assessment does not stop at the technical observation that an AI model contains no folder holding all the files it was once trained on.

The question becomes rather: what has the model actually taken from those works?

The text and data mining exception is not a blank cheque

The European exception for text and data mining also came up.

That exception makes it possible, in certain circumstances, to analyse copyright-protected works by automated means. Its reach in the context of AI training is therefore heavily debated.

The court accepts that text and data mining can in principle be relevant to AI training. It draws a line, however, where the system goes beyond analysing information and protected parts of works are actually taken over into the model.

Such memorisation, the court holds, does not fall within the exception.

That distinction matters. It does not mean that every form of AI training on copyright-protected material is automatically unlawful after this judgment. What the court does make clear is that reliance on text and data mining will not suffice on its own once a model turns out to be capable of reproducing specific protected elements of works.

There was a further element in this case. According to the court, Suno had obtained the music in question from YouTube by circumventing a technical measure protecting against downloads. That too meant Suno could not successfully rely on the exception.

Permission remains the starting point

For authors, the significance of the judgment lies mainly here.

Generative AI changes the technology through which creative works are used, but the exclusive rights of authors do not automatically disappear with it. A work can be analysed digitally, converted into tokens and end up in billions of model parameters. That makes the copyright question more complicated, though not necessarily less relevant.

That fits with BumaStemra’s response to the judgment. The Dutch collecting society sees the ruling as confirmation of the principle that authors should retain control over the use of their work for generative AI.

That control has an economic dimension as well.

AI models such as Suno can only function because they learn from large quantities of existing material. Where protected creative works are a relevant raw material, the question naturally arises who benefits from the economic value being created.

GEMA had asked Suno for a licence beforehand. Suno did not take up the offer. At the same time the company exploits its AI model commercially, including through paid subscriptions. The court has now ordered Suno to provide information about the extent of the use, and has held the company liable for damages.

The discussion is shifting as a result, from a question of principle (may copyright-protected material be used for AI at all?) towards a practical follow-up question: on what terms, and for what remuneration?

Not every AI output is therefore an infringement

At the same time, the scope of the judgment should not be overstated.

The court did not hold that all music made with Suno amounts to copyright infringement. Nor does it follow from the judgment that every AI provider needs permission from every individual author for every form of training.

The case concerned six specific compositions and a particular set of technical and factual circumstances. The court found, among other things, that these works had been memorised in the model and that protected elements of them could be discerned in the generated music.

The judgment is also not yet final. Suno may appeal and has already said that it disagrees with the decision.

It is too early, then, to speak of a definitive European answer on the copyright status of AI training.

From opt-out to negotiation

The judgment is nonetheless significant for authors and other rightholders.

Much of the debate about generative AI in recent years has turned on how authors could prevent their work from being used for training. The emphasis lay on technical opt-outs, transparency about training data and the scope for reserving rights.

The GEMA judgment shows a different side of that discussion.

If the use of protected material genuinely falls within the author’s exclusive right, the question is not only how an author can object. Room then also opens up for licences, collective negotiation and remuneration models.

That is a material difference.

For individual authors it is practically impossible to check whether their music formed part of datasets containing millions of files. Collective management organisations such as GEMA and BumaStemra can play an important role there. They can bundle rights, offer licences and, as these proceedings show, act on behalf of large groups of authors where the parties cannot reach agreement between themselves.

The development of generative AI need not therefore stand opposed to the interests of authors. A market may also emerge in which technology companies gain access to high-quality creative material and authors are compensated for it.

For that to happen, though, it must first be clear that creative works are not simply a free raw material.

An important step, but not the end of the road

The Munich judgment does not resolve the wider conflict between generative AI and copyright. The technology, the various training methods and the proceedings currently under way in different countries are too varied for that.

What the ruling does make clearer is one thing.

The fact that a copyright-protected work is processed by a complex AI model does not mean that the author’s position disappears at the model’s door. Where a model takes over protected elements of a work and is then able to reproduce them, copyright can continue to play a role within that technical chain as well.

For authors that is perhaps the most important significance of this case.

The discussion is slowly shifting, from whether their rights are still relevant in the age of generative AI, to how those rights are to be respected and paid for in practice.

And that second question could ultimately prove more important than the answer in this single set of proceedings.