A Value Sensitive Design Perspective on AI Biases

Abstract

Artificial Intelligence (AI) technology has made profound impacts in our society but concerns about AI biases are rising. This paper classifies AI-related biases and proposes strategies to tackle them. To inform our study, we review AI research on human values to identify three categories of AI biases: pre-existing, technical, and emergent. Informed by the value sensitive design (VSD) framework, we then map the AI biases to the three phases (conceptual, empirical, and technical) of VSD investigation. Our analysis shows that both conceptual and empirical investigations are helpful for addressing pre-existing bias, technical investigation for technical bias, and both technical and empirical investigations for emerging bias. The paper highlights that to effectively tackle AI-related biases, it is important for AI developers and the user community to understand human values in an AI context and to advocate for developing AI-specific value-oriented standards that are agreed upon and adopted by all stakeholders.