Sergio Paulo Caetano

AI systems deployed in high-stakes domains such as lending, hiring, and policing can reproduce or amplify inequities through biased data, feedback loops, contested fairness assumptions, and changing institutional and societal conditions. Although prior work offers many techniques for fairness measurement, testing, and mitigation, the literature remains predominantly model-centered and only weakly connected to requirements engineering concerns such as explicit requirement formation, traceability to deployment context, acceptance criteria, monitoring triggers, revision ownership, and governance over time. To address this gap, we conducted a systematic literature review of 110 studies published between 1996 and 2025, following Keele’s software engineering review protocol, and analyzed the literature through a requirements engineering lens across two organizing axes: Environment-System Design-Sociotechnical levels and lifecycle activities from elicitation to revision. The results show that fairness is discussed across all three levels, but explicit and testable fairness requirements are concentrated at the system-design level. Across the lifecycle, the literature is strong in specification, operationalization, and offline validation, but weak in elicitation, monitoring, revision, and governance. The synthesis further shows that fairness requirements often fail to remain valid over time because context assumptions, acceptance criteria, monitoring signals, revision triggers, ownership, and change-control mechanisms are frequently missing or underspecified. In response, we provide a systematic synthesis of the field, a lifecycle characterization of fairness practices, a longitudinal gap analysis, and the Multidimensional Iterative Fairness Framework, thereby repositioning fairness as a lifecycle-oriented requirements engineering problem and providing a roadmap for future research and practice.