"Are you sure about that?" The effects of calibrated classification model task accuracy and confidence on...

Artificial intelligence and machine learning are becoming increasingly popular. With their deployment comes the challenge of ensuring people have appropriate expectations of their function to ensure calibrated reliance. As these models are traditionally opaque, emerging models are attempting to better convey their confidence and accuracy to users. The present research expands on recent work, investigating the effects of models' classification confidence and accuracy on trust-relevant criteria. In a within-subjects design, participants leveraged models of varying confidence and accuracy in an online image classification task. Results demonstrated the effects of model classification accuracy were qualified by its classification confidence. Post hoc analyses showed that in most cases, models conveying high confidence and low accuracy were perceived as less trustworthy, trusted less, and resulted in lower task performance and increased decision time. The results replicate previous work and expand on the implications that model confidence and accuracy have on trust-relevant criteria.