AI-Competent Individuals and Laypeople Tend to Oppose Facial Analysis AI

Abstract

Recent advances in computer vision analysis have led to a debate about the kinds of conclusions artificial intelligence (AI) should make about people based on their faces. Some scholars have argued for supposedly "common sense"facial inferences that can be reliably drawn from faces using AI. Other scholars have raised concerns about an automated version of "physiognomic practices"that facial analysis AI could entail. We contribute to this multidisciplinary discussion by exploring how individuals with AI competence and laypeople evaluate facial analysis AI inference-making. Ethical considerations of both groups should inform the design of ethical computer vision AI. In a two-scenario vignette study, we explore how ethical evaluations of both groups differ across a low-stake advertisement and a high-stake hiring context. Next to a statistical analysis of AI inference ratings, we apply a mixed methods approach to evaluate the justification themes identified by a qualitative content analysis of participants' 2768 justifications. We find that people with AI competence (N=122) and laypeople (N=122; validation N=102) share many ethical perceptions about facial analysis AI. The application context has an effect on how AI inference-making from faces is perceived. While differences in AI competence did not have an effect on inference ratings, specific differences were observable for the ethical justifications. A validation laypeople dataset confirms these results. Our work offers a participatory AI ethics approach to the ongoing policy discussions on the normative dimensions and implications of computer vision AI. Our research seeks to inform, challenge, and complement conceptual and theoretical perspectives on computer vision AI ethics.