Facial Recognition Technology and Racial Discrimination a Study on the Design of On-Demand-Asynchronous Video Interview Solutions Used in Recruitment

Abstract

7% of people of colour in the UK are unemployed as compared to 4% for whites (UK Gov, 2021). While recent breakthroughs in face recognition technology (FRT) have enhanced the process of recruitment through the use of on-demand- asynchronous video interview solutions (Tambe, et al., 2019; Harwell, 2019; Nawaz, 2020), several other studies have however confirmed that FRT and its application in recruitment activities are biased towards people of colour (Izario et al., 2017; Buolamwini, 2019; Simonite, 2019). Consequently, this is predicted to lead to an increased diversity crisis within the workplace as the employment of people of colour may even lessen further (West et al., 2019). The aim of this dissertation was therefore to understand how on-demand-asynchronous video interview solutions (ODAVIS) capture peculiarities in faces. Specifically, how creators of FRT determine candidates’ suitability for jobs using FRT’s facial analysis, if the data used to train their model is representative of the faces in the world and what reasonable adjustments (if any) are made for people of colour. The major findings include the use of facial landmarks and gaze estimators to analyse candidates’ facial expressions and emotions during interviews, the use of non-representative training data that leads to inconsistency in performance, and the inclusion of certain adjustments (like automatic flashlights, data augmentation or bias audit) during the designs of the solutions and implementation, to close the gaps of inconsistent performance observed for different races.