Abstract

Efforts to address algorithmic harms have gathered particular steam over the last few years. One area of proposed opportunity is the notion of an “algorithmic audit,” specifically an “internal audit,” a process in which a system’s developers evaluate its construction and likely consequences. These processes are broadly endorsed in theory—but how do they work in practice? In this paper, we conduct not only an audit but an autoethnography of our experiences doing so. Exploring the history and legacy of a facial recognition dataset, we find paradigmatic examples of algorithmic injustices. But we also find that the process of discovery is interwoven with questions of affect and infrastructural brittleness that internal audit processes fail to articulate. For auditing to not only address existing harms but avoid producing new ones in turn, we argue that these processes must attend to the “mess” of engaging with algorithmic systems in practice. Doing so not only reduces the risks of audit processes but—through a more nuanced consideration of the emotive parts of that mess—may enhance the benefits of a form of governance premised entirely on altering future practices.