Literature Review: Mechanistic Foundations of Algorithmic Disparities
Contemporary theoretical critiques of automated evaluation demonstrate that the standard practice of withholding protected attributes is structurally insufficient for ensuring equitable decisions. Advanced predictive architectures systematically reconstruct demographic traits through latent correlation structures present across ostensibly neutral input features [4]. In algorithmic credit evaluation, geographic variables, professional categorization, and commercial network structures serve as dependable proxies for sensitive socio-demographic indicators [1], [4]. Consequently, an artificial intelligence system trained on historical portfolios reproduces established patterns of structural stratification under the guise of statistical objectivity. The conceptual mechanics of this phenomenon diverge considerably across distinct model families. Standard linear models maintain uniform sensitivity along predictable functional planes, whereas non-linear estimators, such as gradient-boosted decision trees and artificial neural networks, capture complex high-order feature interactions that amplify subtle historical disparities [2]. Comparative research emphasizes that model selection profoundly influences the manifestation of bias, with certain ensemble architectures exhibiting greater stability and lower baseline discriminatory divergence than unconstrained neural networks [2]. As predictive algorithms exploit these non-linear interdependencies, audit frameworks must transcend simple input verification, focusing instead on internal representation mapping and post-hoc attribution protocols to detect discriminatory dynamics.