Raising User Awareness of Bias-Leakage via Proxies in AI Models to Improve Fairness in Decision-making

Artificial Intelligence systems are becoming more common in decision-making, both for facilitating automated decisions or in tandem with human decision-makers as decision-support systems. AI-assisted DSS are typically employed to make data-driven recommendations to human decision-makers in an effort to improve efficiency and accuracy. In addition, the AI used to power DSS are typically blackbox in nature, meaning that human decision-makers are unaware of exactly how these systems are coming to their conclusions. This is problematic since research in algorithmic fairness already shows that data-driven AI systems can be influenced by social biases present in training data, to reinforce systemic biases and perpetuate unfairness towards minority social groups. When used in high-stakes decision-making, such systems risk protracting systemic biases and further driving social division. An area of research is emerging acknowledging that unfairness can leak through 'proxy' features, causing an implicit-bias effect. In this work-in-progress paper, we propose explaining fairness properties of AI systems and their downstream social impacts to decision-makers - by visualising bias-leakage through proxies - for improved fairness. Finally, we are currently in the process of conducting a study to empirically assess how visualising proxy-biases in AI-assisted DSS can affect decision-making and improve fairness.