Fairness Beyond Prediction: Rethinking Alignment Procedures for Agentic AI Systems

Kaustubha V

Track: Tiny/Short Papers Track (up to 3 pages)

Keywords: algorithmic fairness, AI alignment, agentic systems, procedural fairness, long-horizon impacts, governance

TL;DR: Fairness in agentic AI systems must be evaluated across alignment procedures, not just model outputs.

Abstract: As AI systems evolve from static predictive models to interactive, agentic systems that plan, adapt, and act over time, classical algorithmic fairness frameworks become insufficient. Traditional fairness notions were largely developed for single-shot prediction or decision-making and fail to capture procedural, temporal, and alignment-induced disparities that arise in modern deployments. This paper argues that fairness must be understood as a property of alignment procedures rather than model outputs alone. We identify key fairness failure modes introduced by value learning, policy adaptation, and long-horizon agent behavior, including feedback-driven amplification and path-dependent disparities. We propose a lightweight conceptual framework for analyzing fairness across alignment stages and discuss implications for evaluation, auditing, and governance of agentic AI systems.

Submission Number: 3