Transparency and trustworthiness: Exploring humanmachine interaction in an image classification task

Human interaction with machine learning (ML) is gaining traction as both a research topic and potential technology force-multiplier. As such, identifying which and when ML products are perceived as trustworthy, are relied upon, and affect human performance is crucial. Recently, software engineers have developed calibrated classification models (CCMs) – i.e., ML products which provide more accurate confidence bounds of their performance to users than traditional, uncalibrated classification models (UCMs). The current study investigated whether UCMs and CCMs had different effects on trustworthiness, trust, and human performance in an image classification task. In addition, the relevance of antecedents to trust toward UCMs and CCMs using a new, appropriately constructed scale were explored. Participants completed a task in which they classified images as cats, dogs, or other mammals while a partnering UCM or CCM did the same. Participants had the opportunity to change their partner algorithm's decisions. Results showed that although there were no differences between conditions on criteria, trustworthiness perceptions of both UCMs and CCMs accounted for significant variance in intentions to rely on said algorithms and performance in the task. Implications are discussed, particularly on the relevance of the trust process in human-ML interactions.