Data-Driven Analysis of AI Bias in Public Sector Decision Systems

Decision-making systems in the public sector have become dependent on AI systems for decisions with significant consequences, such as loan approvals in publicly guaranteed lending systems. Although these systems offer efficiency and minimal error margin, they incorporate systematic biases that impact marginalized communities adversely. This research explores an extensive AI bias analysis in the public sector using a hypothetical pipeline for loan approvals. The paper focuses on identifying the three main types of AI bias: (i) data bias caused by historical underrepresentation and proxy discrimination in the training dataset, (ii) model bias due to algorithmic exaggeration of spurious correlation in the training process, and (iii) human bias introduced through feature engineering, decision thresholds, and administrative overriding of the model predictions. The study uses the German Credit Dataset as a surrogate for public-sector lending data. We explore the bias-accuracy trade-off quantitatively before and after pre-processing reweighting. Indeed, the empirical evidence shows that post-mitigation fairness (in terms of DPD and EOD) improves substantially (from 0.28 to 0.06), whereas there is only marginal degradation in predictive accuracy (from 75.4 % to 72.1 %). These results correspond to the existing literature in terms of significant benefits from fairness intervention with a modest price paid in terms of accuracy loss (de Castro Vieira et al. [16]; Chen et al. [13]). To complement the discussion, the paper also considers the simulations of human overrides, thus demonstrating their potential in increasing discrimination. Overall, the research gives a clear example of how AI applications can achieve fair and yet still accurate results, which makes it relevant for the government agencies in terms of policy recommendations. It also offers a clear list of biases and how to mitigate them to achieve fair results at reasonable accuracy costs (Mehrabi et al. [14]; National Institute of Standards and Technology [2]).