Beyond Templates: Understanding and Addressing Human-AI Interaction Harms Through Practitioners' Assumptions

Abstract

Human-AI interaction risks account for most real-world AI harms but remain underrepresented in safety evaluations. Instead, these tend to prioritize model-level evaluations, abstracting away the contexts in which harms emerge. In tackling this, responsible AI efforts have provided practitioners with tools, such as checklists and impact assessments. Yet, these tools often assume a shared understanding of harm, overlooking practitioners' personal, organizational, and media assumptions. As research increasingly addresses human-AI interaction risks, it is crucial to examine practitioners' assumptions. I first conduct a survey and interviews to empirically explore how practitioners envision harm through their underlying assumptions. Second, I reflect on these findings to explore how responsible AI efforts can better support critical reflection on underlying assumptions.