FEATURED ACADEMIC SCHOLARSHIP
External publication — not published by ECDLR.
This page presents an independent ECDLR editorial note on an external academic publication. The original work remains available from its original source.
Authors
Ivo Emanuilov · Thomas Margoni
Affiliation
Centre for IT & IP Law (CiTiP), KU Leuven, Belgium
Original Publication
Emanuilov, Ivo / Margoni, Thomas, Forget Me Not: Memorisation in Generative Sequence Models Trained on Open Source Licensed Code, SSRN, 2024.
The paper was first posted on SSRN in March 2024 and examines the copyright implications of memorisation in generative AI systems trained on open-source software code.
ECDLR Note
Generative AI systems do not merely learn general patterns from their training data. In certain circumstances, they may also memorise and subsequently reproduce fragments of the material on which they were trained. Emanuilov and Margoni examine this phenomenon through a particularly significant example: source code made publicly available under open-source licences.
The paper focuses on situations in which generative models reproduce source code verbatim or in substantially similar form. Such outputs may raise copyright questions, but they may also create additional problems under open-source licensing regimes where attribution, copyright notices or licence information must be preserved.
The authors use the controversy surrounding AI-assisted code generation to demonstrate that the legal assessment of generative AI cannot be limited to the initial training phase. The behaviour of the resulting model and the content generated for users may raise separate questions concerning reproduction, copyright infringement and compliance with applicable licence conditions.
For ECDLR, the study is particularly relevant to the broader debate on the relationship between generative AI, text and data mining, copyright protection and licensing. It also illustrates the growing importance of distinguishing between the legality of obtaining and processing training material and the legal consequences of outputs produced by trained models.
Keywords
Generative AI · Copyright · Open Source · Memorisation · Text and Data Mining · Machine Learning
Original Source
Read full paper on SSRN
View on Zenodo
Featured Scholars
Ivo Emanuilov
Centre for IT & IP Law (CiTiP), KU Leuven
Research focus: AI, copyright, open-source software and technology law
Thomas Margoni
Centre for IT & IP Law (CiTiP), KU Leuven
Research focus: copyright law, digital technologies, AI, data and licensing