George Mason University announced on May 1 that new research led by Jingyuan Yang, assistant professor of information systems and operations management at the Costello College of Business, is exploring how artificial intelligence can promote both fairness and efficiency.
The research addresses concerns that AI systems often reinforce existing inequalities because algorithms learn from data reflecting the current world, which may not be fair or meritocratic. Yang said, “The standard view is that fairness is a tax on efficiency. The way conventional systems are structured, fairness checks are added almost as an afterthought that is assumed to negatively impact system performance.”
Yang’s ongoing work with Pengzhan Guo of Duke Kunshan University and Keli Xiao of Stony Brook University proposes a different approach. Their “fairness-by-design” framework uses reinforcement learning in a dynamic environment where multiple agents compete for limited resources over time. According to Yang, “Our ‘fairness-by-design’ framework utilizes reinforcement learning, which is a type of machine learning (ML). But unlike most machine learning algorithms, ours includes multiple agents competing for finite resources in a dynamic environment, not a static one. That makes our paradigm much more structurally similar to many real-world environments in which various people compete over time for finite resources.”
The framework integrates fairness by encouraging high-performing agents to explore more options while allowing lower-performing agents to settle into stable paths sooner. Additionally, opportunities abandoned by top performers are redistributed first to those who performed less well. As Yang summarized: “The exploratory activity of the high performers releases opportunities that the system channels down toward the weaker performers. Theoretically, this increases fairness while retaining individual choice and without constraining performance.”
To test their ideas, researchers analyzed job histories from 6.5 million professionals across two decades and converted this data into simulated opportunities for hypothetical agents within their algorithmic model. Performance was measured by aggregate rewards earned; fairness was defined as reduction in initial disparities over time.
Their method outperformed eight other machine learning approaches across both metrics—fairness and performance—even when adjusting for changing preferences among individuals at different career stages or applying it to unrelated domains such as taxi route optimization using New York Yellow Taxi Trip records.
Yang said the adaptability shown means the approach could be used as a governance mechanism in future AI applications such as health care scheduling or course registration: “Because the framework proved adaptable to different domains and agent preferences, we think it could be used in future as a governance mechanism for a variety of AI contexts.” She concluded: “Our formal proof establishes the conditions under which fairness and performance reinforce each other, and our experiments show those conditions are achievable in realistic settings. That gives our work both theoretical and experimental grounding.”
