The “ChatGPT” judgment of the Munich I Regional Court (42nd Civil Chamber) marks the first time a German court has affirmed the reproduction of training data within generative AI models and denied the applicability of the text and data mining exception (§ 44b UrhG). Tim W. Dornis and Sebastian Stober analyze the decision from both a technical and a legal perspective, demonstrating that generative AI models are functionally comparable to classical data encoding methods (e.g., JPEG, MP3) by virtue of their deterministic mode of operation. The article clarifies that the apparent non-reproducibility of AI output is attributable to external factors and does not preclude the application of the concept of reproduction in the legal sense. The court's “memorization” concept is subjected to critical scrutiny and refined with respect to incomplete incorporations of training data. On the question of the TDM exception, the authors argue for a teleologically coherent either/or rule: where AI training interferes with exploitation rights, the preparatory data collection activities cannot be justified under § 44b UrhG either.
Citation: Gewerblicher Rechtsschutz und Urheberrecht (GRUR) 2026, issue 9, pp. 620–631.
