Assessing Algorithmic Fairness and Bias in Predictive Social Data Science Models in German Public Administration

Abstract

This study investigates the potential for algorithmic bias in Machine Learning (ML) systems deployed within German public administration, specifically in social resource allocation and urban planning decision support, employing a Computational Fairness Audit methodology that integrates statistical bias measurement with sociologically grounded fairness theory. The research operates within the layered regulatory environment of the European Union's Artificial Intelligence Act, the General Data Protection Regulation, and the German Sozialgesetzbuch, constructing an audit protocol that is simultaneously technically rigorous, legally compliant, and sociologically informed. Using simulated administrative datasets and pseudo-anonymised historical records compiled in compliance with German federal data protection law, the study operationalises six socially grounded fairness metrics applied to three ML model architectures: logistic regression, gradient boosting, and a deep neural network, trained on tasks representative of Jobcenter benefit allocation scoring and municipal social housing prioritisation. Statistical measurement of implicit discrimination against minority and socially vulnerable population groups defined with reference to the protected characteristics enumerated in the Allgemeines Gleichbehandlungsgesetz reveals consistent and statistically significant fairness metric violations across all three model architectures, with gradient boosting and deep neural network models demonstrating substantially higher demographic parity gaps against Turkish-German, refugee-background, and single-parent household population subgroups than against the majority population reference group. The audit further reveals that model accuracy, as measured by standard classification performance metrics, is inversely correlated with fairness metric performance across protected subgroups. This finding directly challenges the technically unwarranted assumption prevalent in German public-sector AI procurement that high model accuracy implies equitable decision outcomes. The study contributes the Socially Grounded Algorithmic Audit Framework (SGAAF) as a replicable, legally anchored protocol for the pre-deployment and in-deployment assessment of ML systems in German public administration contexts.