Achieving Demographic Parity Across Multiple Artificial Intelligence Applications: A new approach for Real-Time Bias Mitigation
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Through quantitative analysis of three datasets, this study examines the efficacy of synthetic data generation in mitigating demographic bias within artificial intelligence (AI) systems across multiple sectors. It evaluates approaches based on Generative Adversarial Networks for creating demographically balanced synthetic data whilst maintaining data fidelity and model performance. The findings demonstrate significant improvements in fairness metrics, with Demographic Parity scores increasing markedly for both race and gender attributes, whilst maintaining comparable accuracy scores between synthetic and original datasets (0.87 versus 0.88). Kullback-Leibler Divergence values below 0.003 for categorical features indicate successful replication of demographic characteristics. The study reveals that synthetic data can effectively address representational imbalances without compromising predictive performance, particularly in bias-sensitive domains such as healthcare and criminal justice. However, implementation requires rigorous validation protocols and human oversight to ensure quality and fairness. The research contributes to the growing body of evidence supporting synthetic data as a viable solution for developing more equitable AI systems, particularly within regulatory frameworks emphasising fairness and privacy. These findings have significant implications for the advancement of AI development, suggesting the integration of human-in-the-loop systems to enhance synthetic data generation quality and fairness.