Algorithmic Impact Assessments at Scale: Practitioners’ Challenges and Needs

Abstract

Algorithmic Impact Assessments (AIAs) are often suggested as a tool to help identify and evaluate actual and potential harms of algorithmic systems. While the existing literature on AIAs provides a valuable foundation, critical understanding gaps remain, including the lived experiences of practitioners who implement assessments and a lack of standardization across industry. Such gaps pose significant risks to the usefulness of assessments in the responsible development of algorithmic systems. By conducting 107 assessments with practitioners who build personalization, recommendation, and subscription systems at a large online audio-streaming platform and 8 semi-structured stakeholder interviews, we attempt to bridge this gap by identifying practitioners’ challenges when applying AIAs that might hinder their effectiveness. The paper analyzes whether harms described in the literature related to machine learning and recommendation systems are similar to the concerns practitioners have. We find that the challenges practitioners encounter can be grouped into three categories: technical and methods, infrastructure and operations, and resourcing and prioritization. We also describe ways for teams to more effectively mitigate concerns. This paper helps bridge gaps between the theory and practice of AIAs, advances understanding of the potential harms of algorithmic systems, and informs assessment practices to serve their intended purpose.