One of the most insidious risks associated with AI is the possibility that its applications yield mixed results—enthusiastic deployments accompanied by harms that remain uncorrected or unevaluated. For example, a tool may enhance capabilities but also introduce risks, stress, or additional burdens to individuals or society. Furthermore, a tool might generally benefit the population while exhibiting bias or marginalizing specific subgroups. An illustrative case is an AI tool designed to improve the efficiency of academic admissions processes, which might inadvertently result in marginalization or unequal resource distribution. Decision support tools may outperform humans on average but often lack the common sense or values required to handle outlier cases in a humane manner.
Actionable recourse refers to an individual's actions to achieve a desired outcome from a machine learning model. Various domains, such as lending, insurance, resource allocation, and hiring decisions, advocate for recourses to ensure trust in the decision system. In these scenarios, it is crucial to ensure the actionability of recourse; otherwise, the suggestions become meaningless. Counterfactuals should adhere to user-defined global and local constraints to generate personalized recourses that recommend actionable changes to the user. However, it is equally important to study the human-AI interaction and how feedback loops impact fairness. This is a vital aspect, as AI could exacerbate inequalities, cause job disruptions, or lead to an undesired distribution of some attributes, such as resources, opportunities, services, etc, over time, even if a single instance of suggested recourse appears reasonable.
In this master’s thesis, we will examine counterfactual actionable recourse and its impact on fairness over time. Our investigation into the interplay between recourse strategies, resource constraints, variable distributions, and their implications on fairness over time contributes to the study of responsible AI.