Copyright lessons on Machine Learning: what impact on algorithmic art?

Abstract

Nowadays, Artificial Intelligence (AI) is described as “the new electricity”. Current algorithmic innovation allowed the development of software which enables machines to learn and to achieve autonomous decision making, with limited or no human involvement, in a vast number of applications, such as speech recognition, machine translation and algorithmic creation of works (computer generated art), on the basis of a process widely known as Machine Learning (ML). Within the ML context, machines are repeatedly trained by means of specifically designed learning algorithms that use a corpus of examples in the form of data sets as training material. Very often and, especially in the context of algorithmic creativity, the training material is mainly composed by copyrighted works, such as texts, images, paintings, musical compositions, and others. Machine Learning workflow typically involves the realization of (multiple) reproductions of any protected work used as training material. The present paper aims to assess the extent to which the use of copyrighted works for Machine Learning purposes in the field of algorithmic creativity is controlled by the monopolistic power of the copyright rightholder on that work. The answer to this question will be researched in the context of EU copyright law, by examining the content of reproduction right and exceptions possibly applicable in a typical ML workflow in the field of algorithmic art, before making an overall assessment of the current EU regulatory framework for artistic ML projects, as it is shaped after the DSM Directive 2019/790.