Copyright Issues in Deep Generative Models According to European Union Law

Over the past decade machine learning witnessed an unprecedented increase in popularity, mainly due to the excellent results achieved by neural networks on a large variety of tasks. Among the subfields of machine learning, Generative Deep Learning stands out for the interest it is lately drawing in the research community and general public. In particular, the last few years marked the dawn of a new era in image and text generation, thanks to new powerful models which are deeply impacting the way we work, communicate and even create art. State-of-the-art text-to-image translation models are now capable of generating novel and photo-realistic images of almost anything, while in the realm of text generative models we are witnessing the overflowing emergence of Large Language Models such as ChatGPT, that are revolutionising the way we work and our daily lives. These developments unveil new vulnerabilities, such as the threat posed by machine-generated art to the work of creative professionals, as well as some relevant legal issues regarding copyright and intellectual property. Firstly, modern generative models are trained on vast datasets, densely populated by copyrighted data in the form of images or texts. Moreover, model owners are usually very opaque when it comes to disclosing details related to training data. Secondly, generative models produce novel data, and it is unclear who can claim copyright protection on them (if anyone), raising an interesting philosophical question: what does it mean for an artwork generated by an algorithm to be original? Machine learning regulations are still largely in the making by European authorities, but in this paper we try to clarify these issues according to the law currently in force in the European Union and from the perspective of the Artificial Intelligence Act.