Explanation Hacking: The perils of algorithmic recourse

At first glance, it might seem simple to give reasons why, but in reality, there are several different kinds of reasons, including a myriad of justifications, that could fulfill this explanatory purpose. These reasons-why serve as explanations. In the context of AI, particularly opaque algorithms, explainable AI techniques are employed to articulate such reasons behind the AI's outputs. One primary research directive on the explainability of AI systems focuses on questions concerning the (non-)epistemic norms required to satisfy our desire-and right-to know why an AI system made its decision. One normative approach that is gaining more and more traction is algorithmic recourse (Ustun et al., 2019; Venkatasubramanian and Alfano, 2020). Recourse explanations give reasons that are actionable and feasible for the end user or data subject to change. If the end user takes on these actionable reasons by implementing changes in their life, then, as the story goes, they would get a decision reversal in the future from the same model, ceteris paribus. Take for example being rejected for a bank loan.