Private Law Theory

ABSTRACT
One of the central concerns of the generative artificial intelligence arcade is whether the training of an AI model runs against copyright laws or not. As a part of the debate, various technological concepts have emerged and served the basis of legal analysis. The present chapter aims to review ‘memorization’ from a copyright perspective. The basic technical features of ‘memorization’ will be presented briefly, together with its separation from similar yet distinct concepts like ‘extraction’, ‘regurgitation’ or ‘compression’. This will be followed by the legal analysis of this term. The chapter takes the view that ‘memorization’ fits into to the right of reproduction under the existing copyright laws of the European Union and the United States. At the same time, the legal consequences of such classification are yet blurry, if significant at all. There are concerns as regards the rules on the quantitative and qualitative features of ‘memorization’, including whether it is more than transitory or trivial, and whether it has any independent relevance, or – as AI developers regularly argue – it is only a ‘rare bug’; the evidencing of ‘memorization’ of training data; the importance of exception or limitation in sidestepping ‘memorization’; or – as regards the European Union (EU) – whether the copyright directives and the AI Act applies to the training of GenAI models carried out outside of the EU. Overall, the chapter takes the view that ‘memorization’ will not play a significant role in court practice, and that the focus seems to be shifted from the liability for the training of GenAI models to the liability for the generation of contents with GenAI systems.

Mezei, Péter, Memorization and Generative AI – A Persistent Issue with Copyright Consequences? (August 5, 2025) in Enrico Bonadio, Péter Mezei and Eduardo Alonso (eds), The Cambridge Handbook of Generative AI and IP in Europe, Cambridge University Press, 2026 (forthcoming).