Accuracy and Explainability in Artificial Intelligence: Unpacking the Terms

Abstract

Artificial intelligence (AI) has permeated many aspects of human life from product recommendations on retailers’ websites to critical decisions affecting healthcare and law enforcement. As such systems become prevalent in high risk areas, explaining their logic to demonstrate issues such as fairness acquire increasing significance. Yet the current focus of machine learning models is the accuracy of decisions rather than their explainability. This paper analyses the findings from two citizens’ juries convened to investigate the perceived trade-off between AI explainability and AI accuracy. While the official juries’ report shows clear preferences for accuracy over explanation in some settings, this paper presents an alternative perspective informed by the concept of ambivalence. By introducing some additional metrics and highlighting the possibilities for different forms of explanation, this research demonstrates how the findings from the citizens’ juries might be otherwise, and the social consequences arising. The paper concludes with some future research directions.