E2-RE4AI: An Ethics-Embedded Requirements Engineering Framework for Lifecycle-Integrated AI Systems

Abstract

In artificial intelligence (AI) systems, fairness, pri-vacy and explainability are seldom expressed as requirements that can be verified. Existing approaches stay at the level of principles, focus on model training, or rely on a single audit after the system is built. We present E²-RE4AI, a requirements engineering (RE) framework that converts ethical principles into measurable ethical non-functional requirements (E-NFRs). Conflicts between these requirements are settled with an analytic hierarchy process whose inputs and result are recorded. The requirements are verified at five gates across the lifecycle and kept under monitoring after deployment; a violation re-opens the design decision it affects. In a synthetic scholarship-admission case, a model that failed the interpretability requirement was ruled out. With weights that put fairness first, the framework selected an equal-opportunity design that narrowed the selection-rate gap from 0.14 to 0.02 without reducing accuracy. Weights that put privacy first, and a legal prohibition on group-specific thresholds, each led to a different selection, and the reasons for both were documented. Across 200 simulated deployments, the monitoring rule produced no false violations and confirmed an injected drift in 73.5% of runs. After a second protected attribute was introduced, thresholds tuned only for income left a gap of 0.107 between the most and least favourably treated subgroups, which intersectional thresholds lowered to 0.073. The framework can be applied; whether it improves outcomes in practice remains to be tested.