Using Artificial Intelligence to classify Jobseekers: The Accuracy-Equity Trade-off

Abstract

AbstractArtificial intelligence (AI) is increasingly popular in the public sector to improve the cost-efficiency of service delivery. One example is AI-based profiling models in public employment services (PES), which predict a jobseeker’s probability of finding work and are used to segment jobseekers in groups. Profiling models hold the potential to improve identification of jobseekers at-risk of becoming long-term unemployed, but also induce discrimination. Using a recently developed AI-based profiling model of the Flemish PES, we assess to what extent AI-based profiling ‘discriminates’ against jobseekers of foreign origin compared to traditional rule-based profiling approaches. At a maximum level of accuracy, jobseekers of foreign origin who ultimately find a job are 2.6 times more likely to be misclassified as ‘high-risk’ jobseekers. We argue that it is critical that policymakers and caseworkers understand the inherent trade-offs of profiling models, and consider the limitations when integrating these models in daily operations. We develop a graphical tool to visualize the accuracy-equity trade-off in order to facilitate policy discussions.