AI blind spots in recruitment, hidden discrimination and quantitative detection methods

Abstract

This article addresses the problem of hidden discrimination in AI‑based recruitment systems. The widespread adoption of automation and artificial intelligence in hiring increases the speed and scale of candidate selection but also contributes to the replication of historical inequalities, a problem that is particularly critical for small and medium‑sized enterprises (SMEs) with limited resources for model auditing. The literature lacks reproducible, applied protocols that are suitable for use in closed SaaS systems and under constrained access to logs. The aim of the study is to develop and empirically validate a practice‑oriented protocol for the quantitative detection and mitigation of algorithmic bias in personnel selection systems that is compatible with SME capabilities. The methodology combines controlled résumé‑pair tests (résumé audit testing), input–output analysis of models, statistical and decision‑tree methods for identifying proxy features, the use of fairness metrics (Disparate Impact, Equal Opportunity Difference, calibration across groups), and iterative corrective procedures (blind preprocessing, feature revision, adjustment of training distributions, and monitoring). Empirical validation demonstrated that applying the proposed set of measures consistently reduces key discrimination metrics with minimal degradation of selection accuracy and throughput; it also identified the most influential proxy features and proposed threshold criteria for triggering corrective actions. The results formalize a step‑by‑step guide for SMEs and a set of simple monitoring tools suitable when access to system logs is limited. The contribution of the study is a practical, reproducible methodology for auditing and correcting AI in recruitment that balances efficiency and fairness and is suitable for implementation under resource constraints. The proposed approach includes recommendations for regular validation, model version documentation, staff training, and integration of explainable AI modules; its adoption reduces legal and reputational risk and increases candidate trust. Further research should evaluate the protocol’s transferability across industries and regions and develop automated tools for proxy‑feature detection.