Copyright and Machine Learning: from human-created input to computer-generated output

Abstract

This thesis examines the legality of unauthorized reproduction of in-copyright works for the purpose of being used in Machine Learning processes. It focuses primarily on US and EU copyright systems as environments for Artificial Intelligence technological developments. Machine Learning uses of creative works differ from traditional ones: they do not involve human readers, they do not display protected expression to the public and they analyse works to extract information not protected by copyright. Hence they raise a question of whether they actually should fall within a reach of exclusive rights of copyright holders. In respect of the US copyright system, this study addresses the fair use doctrine that under certain conditions allows unauthorized reproduction of works. The research makes an attempt to apply the doctrine to Machine Learning uses by drawing parallels with recent case law on other technological uses of copyrighted works. As regards the EU copyright realities, this research discusses Machine Learning uses within the scope of newly proposed copyright exception for Text and Data Mining. It firstly analyses whether exempting these uses from a copyright reach would meet the three-step test requirements. After that, it critically assesses the scopes of the exception proposed and negotiated on the EU policymaking level. Additionally, this study discusses relations between AI-generated works and original human-created works used during the training process. It touches upon a question of possible reproduction of protected expression from original works in secondary ones and copyright-related consequences of that.