Correlating Race: On Machine Learning and Racial Categories

• In Event: Race and/as Technology - II

Abstract

To treat race as technology, as Wendy H. K. Chun suggests, is to acknowledge that race is something that “one uses, even as one is used by it”. To use and to be used in turn neatly describes our relationship with distributed online services, whose free use comes at the cost of the data we produce. This paper will explore what this claim might mean once race is inductively ‘produced’ by machine learning algorithms.

Recent research on the relationship between machine learning and race tends to focus on biases in the data sets used to train particular algorithms, like those used to determine judicial outcomes or those used in facial recognition. We know that we’re incommensurable with the data we produce. Arguably, we need to ask not whether algorithms are racist, but—following Chun—what they use race to do.

Machine learning produces probabilities: instead of distinctions—white/other—it generates percentage matches. By merging a media studies approach to analysing algorithms with science and technology studies research into how categories are produced, this paper will argue that machine learning repurposes the category of race as a correlative. This approach resituates race within the algorithm’s defining binary: opaque/transparent. More crucially, it allows race to be mobilised in the automated production of exclusions. Race enters circulation, subject to new—other—transformations as it’s participatively (re)produced in and as processed data. It’s in circulation, I’ll argue, that we’ll find race’s misuses in the present—and their (political) alternatives.

Author

• Scott Wark, University of Warwick