Recommender Systems are a cornerstone of digital interaction, particularly on platforms such as social media, where personalised content is pivotal to user engagement. One of the main challenges related to such Systems is exposure bias: since users are unconsciously subject to what the System shows them, they engage with a biased sample of items. This proposal introduces the concept of compound exposure bias—i.e., the progressive accumulation and intensification of exposure bias over time, due to Recommender models continuously adapting upon the outcomes of their previous recommendations. Such a recursive cycle may lead to skewed content exposure, thus hindering diverse information access. Additionally, compound bias might contribute to problematic user behaviours, including addiction, since the reliance of many Recommenders on implicit feedback signals inherently equates the time spent on a platform with the degree of user satisfaction with it. We propose to study the dynamics of compound exposure bias through a multi-phase approach, exploring its real-world impact on recommendation accuracy and beyond. Firstly, we will establish formal methodologies to detect and quantify compound bias within the context of recommendation. This involves both theoretical modelling and empirical, longitudinal evaluation. Secondly, we will propose novel strategies to mitigate compound bias while retaining recommendation effectiveness as much as possible, leveraging techniques such as bandits, agentic models, re-ranking, and fairness-constrained learning. This research seeks to enhance the long-term fairness, robustness, and user-centricity of Recommender Systems. Our goal is to offer insights for improving user experience and reducing detrimental effects associated with personalised content consumption.