Private Law Theory

ABSTRACT
This article examines the metaphor of ‘memorisation’ in current debates about artificial intelligence (AI), particularly large language models (LLMs), and its migration into copyright law. In machine learning, memorisation has a narrow technical meaning, describing rare instances where models reproduce verbatim fragments of training data. Yet, when transposed into legal discourse, the term has been misread as equivalent to copyright reproduction. This conflation has far-reaching implications: it risks doctrinal incoherence, undermines the foundations of EU copyright law, and threatens to destabilise the balance carefully struck in the acquis, including the textand-data mining (TDM) exceptions present in Articles 3 and 4 of the EU Copyright in the Digital Single Market Directive. Through an interdisciplinary analysis combining linguistics, computer science, and copyright law, the article argues that the memorisation metaphor is both misleading and legally unsound. Training an LLM does not entail a bit-wise reproduction, storage or the retention of identifiable copies, but rather the optimisation of parameters that enable statistical generalisation. Treating this optimisation as reproduction under copyright law constitutes a category error that expands copyright concepts beyond their intended scope. The article identifies the risks of legislating on the basis of this metaphor: doctrinal drift, policy misalignment, increased compliance costs for small and medium-sized enterprises, barriers to research and open science, and disproportionate benefits for incumbents. It also shows how misplaced remedies based on this misunderstanding risk failing authors while entrenching litigation-driven outcomes. The article proposes constructive alternatives, including models that favour creators, stronger dataset governance, and support for open science. It concludes that regulation must avoid metaphorical shortcuts and instead reflect the technical realities of AI in order to preserve copyright balance, foster innovation, and secure meaningful remuneration for creators.

Larroyed, Aline, The Fallacy of the File: How the Memorisation Metaphor Misguides Copyright Law and Stifles AI Innovation (November 21, 2025).