Systematizing Audit in Algorithmic Recruitment

Abstract

Business psychologists study and assess relevant individual differences, such as intelligence and personality, in the context of work. Such studies have informed the development of artificial intelligence systems (AI) designed to measure individual differences. This has been capitalized on by companies who have developed AI-driven recruitment solutions that include aggregation of appropriate candidates (Hiretual), interviewing through a chatbot (Paradox), video interview assessment (MyInterview), and CV-analysis (Textio), as well as estimation of psychometric characteristics through image-(Traitify) and game-based assessments (HireVue) and video interviews (Cammio). However, driven by concern that such high-impact technology must be used responsibly due to the potential for unfair hiring to result from the algorithms used by these tools, there is an active effort towards proving mechanisms of governance for such automation. In this article, we apply a systematic algorithm audit framework in the context of the ethically critical industry of algorithmic recruitment systems, exploring how audit assessments on AI-driven systems can be used to assure that such systems are being responsibly deployed in a fair and well-governed manner. We outline sources of risk for the use of algorithmic hiring tools, suggest the most appropriate opportunities for audits to take place, recommend ways to measure bias in algorithms, and discuss the transparency of algorithms.

Measuring Personality through Images: Validating a Forced-Choice Image-Based Assessment of the Big Five Personality Traits

“…However, even when group differences in scores are not due to differences in ability, they do not always lead to adverse impact, especially when the analysis is based on a small sample ( EEOC et al 1978 ). Therefore, further validation is needed with a larger sample to more robustly determine whether the reported group differences could result in adverse impact, particularly since the importance of transparency and fairness in the algorithms used in hiring is increasingly a point of concern ( Kazim et al 2021 ; Raghavan et al 2020 ); Mitigating bias: While the potential for adverse impact from this assessment echoes concerns about the fairness of conventional selection assessments ( Hough et al 2001 ), adverse impact associated with algorithmic recruitment processes can be mitigated by removing the items associated with group differences and updating the algorithms ( HireVue 2019 ; Pymetrics 2021 ), unlike with traditional assessments that use a standard scoring key. Further research exploring the potential for mitigating group differences in the algorithms used by this assessment is needed, particularly since there is evidence of measurement bias in the questionnaire-based measure used to construct and validate the algorithms.…”

“…However, even when group differences in scores are not due to differences in ability, they do not always lead to adverse impact, especially when the analysis is based on a small sample ( EEOC et al 1978 ). Therefore, further validation is needed with a larger sample to more robustly determine whether the reported group differences could result in adverse impact, particularly since the importance of transparency and fairness in the algorithms used in hiring is increasingly a point of concern ( Kazim et al 2021 ; Raghavan et al 2020 );…”

Operationalising AI governance through ethics-based auditing: an industry case study

“…The literature on EBA contains few case studies: Buolamwini and Gebru [ 24 ] assessed the efficacy of external audits to address biases in facial recognition systems; Mahajan et al [ 25 ] outlined a procedure to audit AI systems that replicate cognitive tasks in radiology workflows; and Kazim et al [ 26 ] applied a systematic audit to algorithmic recruitment systems. However, there is still little understanding of how organisations implement EBA and what challenges they face in the process.…”

AI audits for assessing design logics and building ethical systems: the case of predictive policing algorithms

“…The extant literature on AI audits emphasises the importance of auditing practices that can take such theoretical foundations into account by systematically assessing not only the technical (e.g., predictive accuracy and explainability) but also the non-technical (e.g., underpinning design logics and principles) [14,15]. This body of work therefore recognises the importance of proactive auditing of AI systems and their design logics as proposed by this paper.…”