Fake people with real influence: Human responses to computer-generated (CG) and GAN faces

Abstract

Recent advances in technology have led people to increasingly opt for online service delivery and social relationships - including with "people" who are not human. Computer-generated (CG) beings are now providing social companionship, romantic partnership, and even psychological and medical advice, which people have easy access to from the comfort of their own home. CG faces are also becoming commonplace in science as face stimuli. Given the growing reliance on CG beings in real life and science, the overarching aim of this thesis was to understand how people respond to CG and hyper-realistic GAN faces relative to human ones. The central finding of this thesis is that people's responses to CG faces are impoverished relative to human ones, except where people cannot tell them apart. Chapter 3 demonstrates that using CG faces instead of human ones in empirical face research (e.g., emotion perception) can lead to different research outcomes. This finding highlights an important issue in the literature: that studies are sometimes using CG instead of human faces as stimuli. Chapter 4 aimed to build on this idea by quantifying how often CG faces are used in empirical face work. This study found the use of CG faces in empirical research has increased since the turn of the century. Importantly, findings showed CG faces are predominantly being used as stand-ins for human face stimuli, to investigate research questions about perception of human faces. Due to growing evidence of different responses to CG and human faces, Chapter 5 meta-analytically synthesised empirical work that directly compared people's responses to CG and human faces across several face processing domains. Results showed: (1) the types of CG faces used until recently are clearly distinguishable from human ones; (2) people's responses to CG faces are poorer than human ones across several face processing domains (e.g., face memory, emotion labelling, first impressions); (3) some responses to CG and human faces do not differ (e.g., looking behaviour, emotion ratings, perceptions of trustworthiness); and (4) there is a concerning lack of data comparing hallmark face processing effects (e.g., inversion effect, N170 response) for CG versus human faces, which has theoretical and practical implications for face researchers. However, recent technological advances have heralded the emergence of new hyper-realistic faces generated by artificial intelligence (i.e., GAN faces). Therefore, Chapter 6 investigated perceptions of humanness and trust in GAN faces compared to human faces. Findings showed people are biased to perceive White GAN faces as human, and do so with high confidence in their incorrect decisions. This bias did not extend to implicit judgements related to having a mind (e.g., curiosity, imaginativeness), although GAN faces were perceived as more trustworthy relative to human ones. Altogether, the present thesis suggests people's responses to the types of CG faces used until recently are impoverished relative to responses to real human faces. However, this thesis comes at the precipice of hyper-realistic GAN faces entering the public domain, raising questions about their potential misuse going forward (e.g., misinformation, revenge pornography).