AI systems are increasingly being deployed in order to automate business, governmental, and institutional decision-making with significant social consequences that remain difficult to predict. Much research on the implications of AI deployment focuses on quantifying the costs and benefits at the population scale.
At a smaller scale, the possibility of long-term, deleterious effects on individuals due to their exposure to the decisions of AI systems at critical junctures in their lives has also been discussed. Consider the scenario in which an individual has had a loan application rejected by an automated AI system. It is essential that this individual not only understands how the AI came to this decision but also has the opportunity to make relevant changes to their application in order to improve their odds of success. Indeed the right to an explanation has been enshrined in the General Data Protection Regulation (GDPR) legislation.
The technical frameworks which address this concern are counterfactual explanations and algorithmic recourse. In this report, we first define counterfactual explanations and extend this concept to the case of uncertain model outcomes. Following this, we discuss algorithmic recourse and develop a more general framework, which we refer to as control-theoretic algorithmic recourse, that models more complex action-based interventional capabilities inspired by the formalism of reinforcement learning.
This paper is part of our tech reports series, in which we share some of the most significant concepts in the causality literature. Our tech reports are written primarily for a technical audience. For a less technical introduction to causal discovery, take a look at our blog.