Private Law Theory

ABSTRACT
This article examines the copyright implications of memorization in AI language models under EU law, prompted by the landmark judgment of the Munich Regional Court in GEMA v OpenAI (LG München I, November 2025). The court held that the ‘memorization’ of copyrighted works in AI model weights constitutes reproduction under Article 2 of Directive 2001/29/EC, and that the subsequent output of memorized fragments amounts to both reproduction and communication to the public. The article critically analyzes this ruling from three perspectives. First, it examines whether memorization – a probabilistic, non-perceptible encoding of information in neural network weights – should be classified as reproduction within the meaning of EU copyright law. While this view prevails in German and international scholarship, a minority position (Leistner and Antoine) argues that memorization falls outside the reproduction right due to the impossibility of identifying specific works within model parameters. Second, the article assesses the applicability of copyright exceptions, particularly the text and data mining exception under Articles 3 and 4 of Directive 2019/790 and the incidental inclusion exception under Article 5(3)(i) of Directive 2001/29/EC. Third, it explores the practical difficulties in applying legal remedies to AI models, given the technical impossibility of selectively removing memorized content. The article concludes that treating memorization as copyright infringement leads to problematic consequences: it effectively undermines the TDM exceptions, privileges closed-source over open-source models, and may discourage the development of European AI infrastructure. Two alternative approaches are proposed: either situating memorization outside the reproduction right altogether, or developing additional criteria-analogous to the CJEU’s approach to communication to the public-that would limit liability to cases of intentional reproduction or failure to implement reasonable safeguards.

Okoń, Zbigniew, Memorization of Copyrighted Works by AI Models Under EU Law (December 10, 2025). Originally published in Polish in Europejski Przegląd Sądowy 2025, no 12, pp 19-25.