IP Law Daily, COPYRIGHT NEWS: German court enjoins OpenAI from reproducing copyrighted song lyrics, (Nov 18, 2025)

By Matthew Hersh, J.D.

The injunction prevents the reproduction of song lyrics within ChatGPT’s language models as well as its outputs.

The owner and operator of ChatGPT committed copyright infringement when it reproduced the lyrics of nine well-known German songs within the large language model that powers the ubiquitous artificial intelligence network, a German federal court has held. The Munich-based tribunal, in enjoining OpenAI from further reproduction of the works, also found that the AI company infringed on the copyrights in those lyrics when it reproduced recognizable portions of them in response to user queries. The public version of the opinion, in German, Case No. 42-0-141139/24, Nov. 11, 2025, can be found here.

The lawsuit was brought by the German music copyright society known commonly by its acronym GEMA. The copyright society acquired the rights to nine German songs, including among others “Atemlos” by Kristina Bach, “Bochum” and “Männer” by Herbert Grönemeyer, and “Über den Wolken” by Reinhard Mey. The rights organization then brought a lawsuit against OpenAI on grounds of copyright infringement and sought entry of a preliminary injunction against the AI company.

Reproduction of the works in training. The court granted the injunction to GEMA, thus barring OpenAI from reproducing the lyrics to the nine songs within its large language models. The court easily found that the works were original and that OpenAI reproduced those works—essentially “memorized” them—within its large language model. The central question before the court was whether OpenAI’s use of the lyrics was protected by an exception under German law for data mining and evaluation. Under Section 44b of the Act on Copyright and Related Rights (a version in English can be found on a German government website here), it is permissible to reproduce lawfully accessible works in order to carry out “text and data mining,” which the law defines as “the automated analysis of individual or several digital or digitised works for the purpose of gathering information, in particular regarding patterns, trends and correlations.” Did that provision protect OpenAI from reproducing the works as part of its training process? The court found that it did not.

The data mining exception did not apply, the court found, because the reproductions OpenAI made of the works were not merely used for training and evaluation purposes, but also were used for the purpose of creating a model that competed with the very creators of those works. “The background to this [provision],” the court emphasized, “is the idea that these reproductions are only created for subsequent analysis purposes and thus do not impair the author's exploitation interests in the work.” Here, the court noted, ChatGPT did not merely evaluate and train on the song lyrics; it also memorized those song lyrics in its database. “The song lyrics as training data were not only evaluated but were completely adopted into the parameters of the model,” the court noted, “which in turn infringes the exploitation interests of the authors.”

A common-sense reading of the data mining provision supported this conclusion, the court noted. According to its recitals, the court emphasized, the data mining exception “should not only promote innovation and new technologies but also protect authors.” In other words, the court noted, the provision “regulates a situation in which the exploitation interests of the authors are not endangered because mere information is extracted and the work as such is precisely not reproduced.” In the case of reproductions in the model, however, the court noted, “the exploitation of the work is sustainably impaired, and the legitimate interests of the right holders are thereby violated. Authors and right holders would thus be left unprotected by an analogous application of the limitation provision, which does not provide for remuneration for the exploitation.”

Nor could OpenAI rely on any other provision of the copyright act. Section 60d of the Act, for example, protects data mining for “scientific research purposes,”—but OpenAI was not a qualifying research organization under that provision, the court reasoned. Nor did OpenAI’s replication constitute “incidental works,” defined under Section 57 of the Act as “works incidental to the actual subject matter being reproduced, distributed or communicated to the public.” Here, the court emphasized, the main purpose of compiling the training datasets “was not the aggregation of copyright-protected content, but the aggregation of datasets that were as comprehensive as possible without any further focus.” OpenAI did not explain, the court emphasized, “why the entire training dataset should be classified as a[n incidental] work within the meaning of the [copyright law].”

The court also found that the reproduction at issue was not justified by the consent of the copyright holders. OpenAI argued that the music publishers made the works “freely accessible to users without restrictions” and therefore, under German law, should be deemed to have authorized any “customary and expected type of use” of the works. But the rights organization adequately showed that it did not make the lyrics available freely to users, and in any event, “the training of models is not to be regarded as a customary and expected type of use that the right holder must expect”—especially as the reproductions of works in the model were seen as undesirable even by OpenAI. “The defendants themselves state that a regurgitation of training data is a rare bug, for the remedy of which continuous work is being done,” the court noted.

Finally, the court found that there was no need to give OpenAI a “grace period” of six months to remedy the issue before entry of an injunction. OpenAI had been on notice of the alleged infringement since a warning letter was sent in November 2024, the court noted. “A period of grace is not to be granted simply because the defendants have been aware of the copyright infringements at least since the warning letter from the plaintiff,” the court emphasized. “Within this time, they could have already trained a new model, which according to the defendants' own statements should take six months.” OpenAI would be enjoined immediately.

Reproduction of the works in outputs. The court separately found that GEMA was entitled to an injunction in connection with OpenAI’s reproduction of the works in its outputs to user queries. To be sure, the court noted, some of the song lyrics were reproduced inaccurately or bolstered by means of “hallucinations.” But that still constituted infringement by OpenAI. “Copyright infringements exist not only in the case of an identical unlawful reproduction of a work,” the court noted, “but also in the case of adaptations or other transformations.” The “decisive criterion as to whether the scope of protection of a reproduction or public communication is opened,” the court emphasized, “is the recognizability of the original elements of the work in the other object Here, the lyrics of some songs were reproduced so extensively, the court noted, that there were “no doubts about the recognizability of the original lyrics in the outputs.”

Nor could OpenAI point to the AI model itself, as opposed to human conduct, as the source of the infringement. OpenAI argued that insofar as ChatGPT’s output of lyrics was generated by a machine, it was an independent (or “double”) creation by that machine, rather than infringement by a human. The court rejected this argument. “It is undisputed that the disputed song lyrics were used as training data for the models,” the court found. “They are also reproduced in the model. If they are reproduced as output due to the input of simply phrased prompts, as in the present case, there is no independent double creation, but the models reproduce the memorized texts.”

Equally importantly, the court noted, copyright infringement in connection with OpenAI’s outputs could not be wholly attributed to OpenAI’s users. “The copyright infringements caused by the outputs are attributable to systemic causes,” the court noted, “the content design of the outputs is attributable to the defendants as operators of the models.” Indeed, the court noted, OpenAI “used the disputed song lyrics as training data for their models, prepared the training set, and carried out the training with the result that the training data was memorized.” OpenAI also acted, the court emphasized, “in full knowledge of the consequences” of its conduct. It would not be able to pawn off the infringement to its users.

Right to personality. But while OpenAI would be enjoined for its copyright infringement, it would not be enjoined for violating German law known “the general right to personality.” The basis for this claim was that OpenAI often caused “deviations from the correctly memorized text and thus hallucinations by randomization mechanisms at the decoding level”—thus taking the rights holders’ work and “distorting it.” But this claim would not succeed, the German court found.

The claim would not succeed, the court emphasized, because the OpenAI outputs were attributed correctly to the song authors. A claim based on an infringement of the general right of personality comes into consideration, the court noted, if, in the presence of a design not created by the author, identity confusions could arise due to the untruthful attribution of authorship to this design. Here, the court noted, the disputed outputs constitute only “transformations and adaptations” within the meaning of the copyright law, which was different from false attribution of authorship. Moreover, the court noted, the naming of the lyricists as authors in the case of transformations and adaptations of their works concerned only their professional context and thus their social, not their private sphere. “The protection of personality in the social sphere is limited and restricted only to certain manifestations,” the court noted, such as “stigmatization or social exclusion.” Such serious effects were not given by the disputed outputs, the court concluded.

News: AINews Copyright TechnologyInternet GCNNews