Attitudes Toward Facial Analysis AI: A Cross-National Study Comparing Argentina, Kenya, Japan, and the USA

Computer vision AI systems present one of the most radical technical transformations of our time. Such systems are given unparalleled epistemic power to impose meaning on visual data, despite their inherent semantic ambiguity. This epistemic power is particularly evident in computer vision AI that interprets the meaning of human faces. The goal of this work is to empirically document laypeople's perceptions of the epistemic and ethical complexity of computer vision AI through a large-scale qualitative study with participants in Argentina, Japan, Kenya, and the USA (N=4,468). We developed a vignette scenario about a fictitious company that analyzes people's portraits using computer vision AI to make a variety of inferences about people based on their faces. For each inference that the fictitious company draws (e.g., age, skin color, intelligence), we ask participants from all countries to reason about how they evaluate computer vision AI inference-making. In a series of workshops, we collaborated as a multinational research team to develop a codebook that captures people's different justifications of facial analysis AI inferences to create a comprehensive justification portfolio. Our study reveals similarities in justification patterns, but also significant intra-country and inter-country diversity in response to different facial inferences. For example, participants from Argentina, Japan, Kenya, and the USA vastly disagree over the reasonableness of AI classifications such as beautiful or skin color. They tend to agree in their opposition to AI-drawn inferences intelligence and trustworthiness. Adding much-needed non-Western perspectives to debates on computer vision ethics, our results suggest that, contrary to popular justifications for facial classification technologies, there is no such thing as a "common sense"facial classification that accords simply with a general, homogeneous "human intuition."