Multi-task adversarial learning detects intersectional algorithmic bias in AI recruitment systems

With the widespread application of artificial intelligence in recruitment, algorithmic bias issues have become increasingly prominent, seriously threatening social fairness and job seekers' rights. Addressing the limitations of existing bias detection methods in detection granularity, fairness assessment, and interpretability, this study proposes a deep learning-based algorithmic bias detection framework. The framework designs a multi-task adversarial learning architecture that achieves fine-grained intersectional bias detection through attention mechanisms; constructs a multi-dimensional evaluation system comprising nine metrics across three major categories including group fairness, individual fairness, and causal fairness, employing Pareto frontier analysis to reveal fairness trade-off relationships; and integrates an interpretability module based on SHAP values, gradient-weighted class activation mapping, and causal mediation analysis to achieve precise localization of bias sources. Experimental results demonstrate that the framework improves intersectional bias detection accuracy by 12-18% points compared to traditional methods, providing effective technical support and theoretical guidance for fairness improvement in AI recruitment systems.