Invisible Feedback Loops: Detecting Passive Bias in User-Facing ML Models

Machine Learning (ML) models integrated into user-facing systems are extremely well-regarded for their ability to automate and personalize experiences. But lying beneath the surface is a nefarious problem: the growth of silent feedback loops. These loops, formed when model outputs quietly influence user behavior, can in turn perpetuate existing model assumptions, leading to passive bias over time. In this paper, we propose an end-to-end system to detect, analyze, and mitigate passive bias due to such feedback loops. We introduce a feedback-aware monitoring system architecture, describe real-world application scenarios, and provide empirical methods to quantify bias propagation. Our approach highlights the performance and ethical consequences of neglecting latent model feedback and suggests deployment guidelines for responsible deployment.