AI-DRIVEN RECRUITMENT: UNCOVERING AND MITIGATING LATENT BIAS IN RESUME SCREENING ALGORITHMS

Abstract

AI-driven recruitment systems promise efficiency but risk encoding and amplifying latent biases in resume screening. This study investigates how data, features, and model choices introduce disparate outcomes across gender, ethnicity, and socioeconomic proxies. We audit common pipelines using counterfactual testing, subgroup metrics, and representation analysis to reveal hidden bias patterns. We then propose mitigation strategies combining data rebalancing, debiasing embeddings, fairness-aware loss functions, and post hoc calibration. Experimental results on benchmark and real-world datasets show improved equity with minimal accuracy loss. We also discuss governance practices, including audit trails, human-in-the-loop review, and transparent reporting, to ensure accountability and regulatory compliance. The findings provide practical guidance for deploying fairer AI in hiring and highlight open challenges in measuring and mitigating bias at scale. Finally, we outline a reproducible evaluation framework and release tools to support continuous monitoring, enabling organizations to balance performance, diversity goals, and legal obligations throughout the recruitment lifecycle across roles, regions, and time to sustain equitable outcomes consistently.