From Statistical Fairness to Epistemic Fairness: The DBSD Framework for AI Bias Mitigation

Abstract

Yair Oppenheim

Artificial Intelligence (AI), Machine Learning (ML), Large Language Models (LLMs), and data-driven decision systems increasingly influence critical domains including hiring, lending, healthcare, insurance, criminal justice, and digital governance. However, contemporary AI systems frequently reproduce or amplify social inequalities through hidden semantic inference processes that operate beyond explicit discriminatory rules or observable prediction outputs. This article introduces the Deep Bias Systematic Deviation (DBSD) framework, a novel semantic-inferential model that reconceptualizes algorithmic bias as an asymmetric inference-flow phenomenon operating across semantic, behavioral, and networked information structures. Unlike conventional fairness approaches that treat bias primarily as unequal outputs, DBSD models bias as a function of knowledge pressure, semantic alignment, stereotype activation, inference current, and epistemic resistance.

The article reviews five major fairness metrics widely used in AI governance and regulatory auditing: Disparate Impact, Statistical Parity, Equal Opportunity, Equalized Odds, and Calibration Bias. For each metric, the article presents detailed numerical examples illustrating both the biased state and the mitigation process achieved through DBSD-based semantic correction mechanisms. The originality of the article lies in its transition from prediction-level fairness toward epistemic fairness. Instead of correcting only final decisions, DBSD regulates the semantic and inferential processes that generate those decisions. Consequently, the framework provides a scalable and theoretically grounded foundation for future research in AI fairness, semantic auditing, inference governance, and algorithmic accountability.

PDF