Distributional Intersectional Fairness in AI-Supported Job Matching: Evidence on Amplification, Trade-Offs, and Residual Risk

"Despite widespread recognition of intersectionality in discrimination research, most algorithmic fairness evaluations remain limited to single-attribute group comparisons or individual-level prediction errors. This paper introduces a distributional approach to intersectional fairness that evaluates how machine learning models reshape the allocation of outcomes across intersecting social groups. Using large-scale administrative labour-market data from public employment services, we analyse an AI-supported job matching system with 141 occupational classes and more than three million observations. We measure intersectional disparities via Jensen–Shannon divergence and compare observed occupational distributions to model-implied allocations. We show that unconstrained models do not merely reproduce existing labour-market structures but systematically amplify intersectional disparities across occupations. Fairness-aware regularization reduces aggregate disparities but introduces a pronounced trade-off with predictive performance and reallocates distortions across occupations rather than eliminating them. Counterfactual averaging achieves substantial reductions in occupation-level disparities while preserving most predictive performance, may provide a more favourable fairness-performance trade-off in this application setting than directly constraining model outcomes. To identify localized fairness risks, we propose an occupation-level audit mechanism that flags cases with unusually large distributional shifts. Even under strong fairness constraints, a small number of occupations exhibit substantial residual distortions, indicating that fairness interventions compress rather than remove inequality. Our results demonstrate that fairness in large-scale decision-support systems must be evaluated at the level of outcome distributions. More broadly, they highlight the limits of purely technical mitigation and the need to combine fairness-aware modelling with explainability, monitoring, and human oversight in real-world deployment." (Author's abstract, IAB-Doku) ((en))