Trust in Artificial Intelligence: Comparing Trust Processes Between Human and Automated Trustees in Light of Unfair Bias

Automated systems based on artificial intelligence (AI) increasingly support decisions with ethical implications where decision makers need to trust these systems. However, insights regarding trust in automated systems predominantly stem from contexts where the main driver of trust is that systems produce accurate outputs (e.g., alarm systems for monitoring tasks). It remains unclear whether what we know about trust in automated systems translates to application contexts where ethical considerations (e.g., fairness) are crucial in trust development. In personnel selection, as a sample context where ethical considerations are important, we investigate trust processes in light of a trust violation relating to unfair bias and a trust repair intervention. Specifically, participants evaluated preselection outcomes (i.e., sets of preselected applicants) by either a human or an automated system across twelve selection tasks. We additionally varied information regarding imperfection of the human and automated system. In task rounds five through eight, the preselected applicants were predominantly male, thus constituting a trust violation due to potential unfair bias. Before task round nine, participants received an excuse for the biased preselection (i.e., a trust repair intervention). The results of the online study showed that participants have initially less trust in automated systems. Furthermore, the trust violation and the trust repair intervention had weaker effects for the automated system. Those effects were partly stronger when highlighting system imperfection. We conclude that insights from classical areas of automation only partially translate to the many emerging application contexts of such systems where ethical considerations are central to trust processes.

Automated summary

This study investigates trust processes in the context of personnel selection, finding that participants initially have less trust in automated systems and that trust violation and repair intervention have weaker effects for automated systems.