RECURSIVE DISCRIMINATION AMPLIFICATION: A Cross-Domain Empirical Test of Closed-Loop Institutional AI

Recursive Discrimination Amplification (RDA) is a dynamic SignalRupture operator inside System Inheritance Studies that tests whether institutional AI systems merely inherit historical disparity or whether they amplify it through closed-loop interaction with the environment. The paper distinguishes four frequently conflated concepts — inheritance, feedback closure, distributional drift, and amplification — and requires evidence for each stage separately. RDA formalizes the dynamic mechanism through disparity trajectories (Δₜ), an amplification ratio (ARₜ), a feedback-closure index (FCI), and two recursion channels: model recursion (AI decisions alter future labels used for retraining) and institutional recursion (AI decisions alter human or organizational behavior, producing new records that influence later decisions even without model updates). Amplification is defined precisely: ARₜ > 1 with demonstrated feedback contribution. The paper tests this architecture across policing, banking, healthcare, and criminal‑legal risk assessment, grading evidence from conceptual possibility to longitudinal causal amplification. The contribution is major: RDA rejects the universal claim that “AI amplifies discrimination” and replaces it with a conditional, falsifiable dynamic theory. Predictive policing provides strong mathematical and simulation evidence of runaway feedback; banking reveals selective-label feedback but also strong counterevidence where FinTech reduces discrimination; healthcare shows powerful inheritance through biased proxies and simulation evidence of feedback-induced model degradation; criminal‑legal systems show theoretical repeat-use amplification but lack longitudinal causal proof. The cross-domain result is precise: AI can amplify inherited disparities only when historical disparity, consequential interaction, endogenous observation, repeated reuse, and insufficient correction align. RDA also clarifies that drift is not inherently discriminatory — drift becomes harmful only when it interacts with asymmetric feedback. The paper freezes the theoretical architecture into a five-stage sequence (Inheritance → Interaction → Feedback Closure → Drift → Dynamic Outcome) and establishes strict falsification boundaries to prevent overclaiming. Its final contribution is conceptual clarity: AI does not simply inherit history; it can become part of the mechanism that produces the future evidence that justifies its next decision. This transforms RDA from a fairness critique into a dynamic SR operator capable of explaining attenuation, persistence, or amplification across institutional loops.