“Memorization” as an act of reproduction: the limits to AI training begin

In this forum, Eduardo del Río Dutú, lawyer in Bardají & Honrado, addresses a European ruling that could mark a before and after regarding the training of artificial intelligence models with protected works. And training without authorization is no longer a gray area.

In recent weeks, a new European milestone has been added to the debate between Intellectual Property and Artificial Intelligence: the judgment of the Landgericht München I (Munich Regional Court I) of 11 November 2025, case 42 O 14139/24 (GEMA v. OpenAI). In this matter, the court concluded that, when a model “memorizes” copyrighted lyrics and can reproduce them (either partially or with variations) from prompts simple, that fixation in the model is equivalent to a act of reproduction, relevant for copyright purposes, also attributing responsibility to the system operator.

This approach contrasts with the US line that, in recent resolutions such as the “Anthropic” case, has begun to differentiate between training with legitimately obtained copies and illicit uses, opening the door to defenses of fair use. In Europe, however, the Munich ruling points to a stricter interpretation, by denying that they cover the “memorization” that allows protected works to be reconstructed through outputs, thus reinforcing the pressure towards training licenses and duties of diligence for developers.

---

“Memorization” as an act of reproduction

The main argument of the plaintiff management entity (GEMA) focused on the fact that the use of song lyrics to train generative AI models implied a infringement of the copyrights of those represented. The court has accepted this argument, confirming that the “memorization” of linguistic works in an AI model is equivalent to a material fixation and, consequently, their reproduction.

The court confirms that the “memorization” of linguistic works in an AI model is equivalent to a material fixation and, consequently, their reproduction.

According to the ruling, even if the work is not stored as a traditional text file, it is integrated into the model structure in such a way that it can be “perceived indirectly” through the results (outputs) generated by the system. This interpretation is very relevant, as it equates the internal and technical process of AI training with the traditional concept of “reproduction.” This discards the idea that the data simply they “traverse” the model to extract abstract patterns, since the court considers that there is a fixation that allows a subsequent exteriorization of the work, whether total or partial.

---

The text and data mining exception, and its limits in AI training

Against the arguments of the management entity, the defendant entities alleged that their activity was covered by legal exceptions, specifically they understood that reproductions are covered by the text and data mining exceptions, which would cover reproductions made to train AI models.

“Memorization” constitutes a autonomous reproduction, which allows the reappearance of the works in the outputs of the system and would directly affect the exploitation interests of the authors.

However, the German court clearly distinguishes between data set creation phase and the AI model training phase. While the first could be considered a purely instrumental reproduction and, therefore, covered by the data mining exception, the second would imply a full incorporation of song lyrics into model parameters. This “memorization” thus constitutes a autonomous reproduction, which allows the reappearance of the works in the outputs of the system and would directly affect the exploitation interests of the authors.

Starting from this premise, the court understands that said memorization was effectively proven by comparing the original letters with the responses generated by the AI, which was capable of play specific fragments of the songs when faced with requests as simple as “What is the letter of […]?” or “Please tell me the second verse too.”

The ruling concludes that the exception does not cover the act of “feeding” the model with the protected works. This reasoning closes a potential escape route for AI developers.

---

The responsibility of the AI ​​model operator

Another of the relevant pronouncements of the sentence is the direct attribution of responsibility to the operators of AI models. In this way, the court does not analyze the authorship of the result generated by the AI, but rather points directly to who develops, controls and commercially exploits the AI ​​model, concluding that the operators are responsible for the infractions that materialize through the outputs of their systems.

The court does not go into analyzing the authorship of the result generated by the AI, but rather points directly to who develops, controls and commercially exploits the AI ​​model.

Taking into account the responsibility of the operators, the ruling establishes the following consequences: the cessation of activity, ordering the defendants to stop reproducing the unauthorized works; the accountability, obliging the defendants to provide detailed information on the extent of use of the works and the income obtained, and compensation, and the rights holders must be compensated for the damages caused.

---

What could this statement mean?

The Munich ruling marks a before and after in the debate on training artificial intelligence models with protected works. From now on, it is much more difficult to maintain that these uses remain outside of copyright or that they can simply rely on legal exceptions designed for other purposes.

In practice, this forces AI operators to make decisions. They either review and clean their data sets, ensuring that they have sufficient rights to train their models, or they assume a increasing legal risk. The court's message is clear: training without authorization is no longer a gray area, but a terrain that can lead to claims and liabilities for copyright infringements.

For rights holders, the resolution reinforces a negotiating position that until now was weak. The door opens to a more structured market for training licenses, in which the use of protected works ceases to be a fait accompli and becomes a contractual element.

The court's message is clear: training without authorization is no longer a gray area, but a terrain that can lead to claims and liabilities for copyright infringements.

This jurisprudential turn also coincides with the advance of first draft of the Code of Practice on transparency of content generated by AI, promoted by the European Commission in development of article 50 of the AI ​​Regulation. The Code anticipates machine-readable marking and labeling obligations for certain AI-generated content, reinforcing the transparency requirements that will soon be fully applicable to suppliers and professional users.

Considered together, the Munich ruling and the progress of this regulatory work point towards a evolution of the legal framework applicable to AI in which training, generation and dissemination of content will progressively cease to be located in areas of indefinition. Everything indicates that the development and exploitation of AI systems will tend to be increasingly conditioned by explicit legal limits, with greater demands for diligence and responsibility for the operators who control and benefit from these technologies.

Eduardo del Río Dutú

Lawyer in Bardají & Honrado