User discrimination of content produced by generative adversarial networks

Abstract

Artificial Intelligence (AI) is increasingly being introduced in several domains for classification and clustering of different types of existing information (e.g., text, images, audio, and video). Recently, improvements to Machine Learning (ML) and new approaches to the design and use of Neural Networks (NNs) enabled the development of algorithms that generate new content that mimics the features of the training dataset. Specifically, Generative Adversarial Networks (GANs) are particularly effective in producing content with unprecedented levels of fidelity. As a result, they can generate realistic images of people, vehicles, and nature. In this paper, we discuss the results of a study that investigated user perception of pictures generated using GANs, with specific regard to portraits featuring faces. Specifically, our experiment involved 551 participants who were asked to classify over 7000 real images and pictures generated by ML algorithms. Our findings show that users show low accuracy in discriminating images and, thus, demonstrate the effectiveness of GANs in producing content that can be perceived as realistic.