Enhancing Fairness, Justice and Accuracy of Hybrid Human-AI Decisions by Shifting Epistemological Stances

From applications in automating credit to aiding judges in presiding over cases of recidivism, deep-learning powered AI systems are becoming embedded in high-stakes decision-making processes as either primary decision-makers or supportive assistants to humans in a hybrid decision-making context, with the aim of improving the quality of decisions. However, the criteria currently used to assess a system’s ability to improve hybrid decisions is driven by a utilitarian desire to optimise accuracy through a phenomenon known as ‘complementary performance’. This desire puts the design of hybrid decision-making at odds with critical subjective concepts that affect the perception and acceptance of decisions, such as fairness. Fairness as a subjective notion often has a competitive relationship with accuracy and as such, driving complementary behaviour with a utilitarian belief risks driving unfairness in decisions. It is our position that shifting epistemological stances taken in the research and design of human-AI environments is necessary to incorporate the relationship between fairness and accuracy into the notion of ‘complementary behaviour’, in order to observe ‘enhanced’ hybrid human-AI decisions.