Algorithmic loafing and mitigation strategies in Human-AI teams

Abstract

Exercising social loafing – exerting minimal effort by an individual in a group setting – in human-machine teams could critically degrade performance, especially in high-stakes domains where human judgement is essential. Akin to social loafing in human interaction, algorithmic loafing may occur when humans mindlessly adhere to machine recommendations due to reluctance to engage analytically with AI recommendations and explanations. We consider how algorithmic loafing could emerge and how to mitigate it. Specifically, we posit that algorithmic loafing can be induced through repeated encounters with correct decisions from the AI and transparency may combat it. As a form of transparency, explanation is offered for reasons that include justification, control, and discovery. However, algorithmic loafing is further reinforced by the perceived competence that an explanation provides. In this work, we explored these ideas via human subject experiments (n = 239). We also study how improving decision transparency through validation by an external human approver affects performance. Using eight experimental conditions in a high-stakes criminal justice context, we find that decision accuracy is typically unaffected by multiple forms of transparency but there is a significant difference in performance when the machine errs. Participants who saw explanations alone are better at overriding incorrect decisions; however, those under induced algorithmic loafing exhibit poor performance with variation in decision time. We conclude with recommendations on curtailing algorithmic loafing and achieving social facilitation, where task visibility motivates individuals to perform better.