AI auditing through performance appraisals: A practice-informed approach

Abstract

AI auditing is projected to be a crucial governance mechanism for ensuring fair and responsible use of AI systems in society. Yet, despite a proliferation on the market of well-intended new AI auditing tools and proposed evaluation frameworks, AI auditing as a practice in the field is still at a formative stage. In this article, we draw on our experience working with two supervisory authorities in the Netherlands to show that there is an urgent felt need in practice for robust, periodic documentation procedures around AI systems that utilize and align with existing audit-enabling processes like maintaining personnel files. Recognizing the practical expertise of existing supervisory bodies tasked with auditing responsibilities, our aim is to contribute to the existing literature on AI auditing by proposing performance appraisal instruments for AI systems akin to those used to appraise employees. We argue how leveraging existing knowledge, know-how and infrastructure around periodically evaluating employees offers a strong foundation for effectively monitoring, evaluating, and documenting the functioning of AI systems in a socio-organizational context over time. We call on the broader responsible AI community to critically engage with the two core ideas presented in this article, viz. the viability of 1) designing instruments which facilitate a rich kind of documentation akin to performance appraisals of employees, and 2) utilizing existing infrastructure and know-how around performance appraisals to facilitate the uptake of such instruments.