Abstract
Following the US case (Bartz et al. v. Anthropic PBC, US District Court, N.D. California, No. C 24-05417 WHA) and the UK case (Getty Images v. Stability AI), it was inevitable that the collision between generative artificial intelligence and copyright would also occur in the European Union.
The clash, which—at least for the time being—has dealt AI a bruising blow, took place in Germany, where the Munich Regional Court I handed down the first European decision on copyright and generative AI, finding OpenAI liable for the unauthorised use of song lyrics during the training of its models.
Without expressing a view here on whether the ruling is persuasive, the decision proceeds from a proposition as simple as it is far-reaching (given the unequivocal terms in which it is stated): for a work to be memorised, reproduced or made available—even by an algorithm—the author’s permission is required.
The case: 9 songs used—without prior authorisation—to train ChatGPT
GEMA—a German collecting society comparable to Italy’s SIAE—represents and administers the rights of more than 95,000 authors and music publishers. On 13 November 2024, it brought proceedings against OpenAI before the Munich Regional Court I. The allegation is clear-cut: the US giant allegedly used protected song lyrics to train large language models (LLMs), specifically those used by ChatGPT, without obtaining the necessary licences.
According to the claimant, the point is clear: the chatbot’s outputs themselves provide evidence of the large-scale use of copyright-protected works in the training dataset, because several works from the repertoire it administers are allegedly recognisable in those outputs—more or less in full and on repeated occasions—including nine fairly well-known songs, such as “Mambo No. 5” by Lou Bega, “Forever Young” by Alphaville, the intro to “Daddy Cool” by Frank Farian, and “Cheri Cheri Lady” by Modern Talking.
According to the claimant, if confirmed, this would reveal conduct directly contrary to the formal reservation of rights made by GEMA on behalf of its members and would therefore amount to a significant copyright infringement.
Unfortunately for OpenAI and the more libertarian proponents of AI, the German court accepted this line of argument.
Memorisation, reproduction and outputs: the cornerstones of the decision
The Bavarian court held that:
- the presence in the outputs of text identical or nearly identical to the protected lyrics indicates that copies of the works must necessarily have been reproduced in some way within the model;
- it is not for the court to determine how memorisation occurs in technical terms: whether it involves copies in the datasets, caching or “parameterisation” makes no difference;
- what matters is that the lyrics are contained in the model in reproducible form and can be retrieved (intentionally or otherwise) during interaction with the user.
In plain terms: if there is a recognisable reproduction of a protected text, regardless of the means employed, it must be authorised. Technology cannot alter this axiom.
OpenAI’s liability
The German court was unequivocal in holding OpenAI fully liable for the infringement. It is the company—and not the users—that does all of the following:
- determines the model’s architecture,
- selects the “base” training data,
- controls how those data are stored and managed,
- determines the reproducibility of the outputs.
Against that background, the defendant’s argument—which invoked the “purely technical” nature of training—was deemed irrelevant.
The court also rejected the argument that the lyrics were, in any event, “freely available online”, because the fact that they are available does not automatically make that availability lawful or permit their reuse.
Why the Text and Data Mining (TDM) exception does not apply
The Munich Regional Court I also clarified that the EU rules on TDM—which are central to AI developers—could not come to OpenAI’s aid for two reasons:
- this is not merely a matter of extracting information, but of reproducing entire works.
- the TDM exceptions cease to apply when the author has expressly reserved their rights. For many of the works administered by GEMA, that reservation had already been made.
In summary, training an LLM cannot automatically be equated with “fair use” in the context of Text and Data Mining.
The “moral” of the ruling
In the view of the German judges, the reproduction in a chatbot’s outputs of recognisable portions of protected works cannot be regarded as a transformative use.
And here, in the writer’s view, we return to the “keyword” whenever we have to assess whether plagiarism or, more generally, unlawful exploitation has occurred: RECOGNISABILITY. If original expressive elements (in the technical legal sense) of the earlier work are recognisable in the later work, alarm bells should ring.
The German judges made precisely this point: the fact that the original elements of the lyrics (used for training) are recognisable is sufficient to constitute an unauthorised act of making available to the public.
OpenAI had tried to shift responsibility onto users. The court rejected that line of argument as well: responsibility remains with the party that designs and controls the model.
“If you need components to build something, you buy them; you do not use someone else’s property.” With these words, Elke Schwager, the presiding judge of the Chamber, sought to sum up the significance of the ruling.
The potential consequences of the ruling
The judicial outcome of the dispute (in brief, the court ordered OpenAI to stop memorising the protected lyrics, prevent their reproduction in chatbot outputs, and provide information on the datasets, how they were used and the revenues generated) is of relatively little interest to us, not least because an appeal by OpenAI appears all but certain, meaning that the case will continue to be discussed.
Of real interest are the potential consequences that would follow if the case at issue were not to remain isolated but instead became the leading case for a common EU judicial approach.
If consent is required to use a work to train an AI model, that consent has economic value. This gives rise to the increasingly plausible prospect of large-scale licensing agreements between technology platforms and rights holders, with negotiated solutions designed to open up (and “guarantee”) new creative opportunities through AI, while ensuring, one hopes, adequate remuneration throughout the value chain.
The GEMA v. OpenAI ruling appears to seek to curb indiscriminate AI training in pursuit of a (fair?) balance between human creativity and algorithmic reworking: a balance in which—even when technological progress is invoked as a shield—the central importance of consent and remuneration can no longer be ignored.
The game has only just begun.
Data di pubblicazione: 4 Dicembre 2025
© Canella Camaiora S.t.A. S.r.l. - Tutti i diritti riservati.
È consentita la riproduzione testuale dell’articolo, anche a fini commerciali, nei limiti del 15% della sua totalità a condizione che venga indicata chiaramente la fonte. In caso di riproduzione online, deve essere inserito un link all’articolo originale. La riproduzione o la parafrasi non autorizzata e senza indicazione della fonte sarà perseguita legalmente.
Daniele Camaiora
Senior Partner dello studio legale Canella Camaiora, iscritto all’Ordine degli Avvocati di Milano e Cassazionista, appassionato di Nuove Tecnologie, Cinema e Street Art.