It's a disaster! : factors affecting trust development and repair following agent task failure

The effects of anthropomorphism, transparency and system failure on task performance and trust toward an autonomous system during human-autonomy teaming (HAT) were investigated. Participants (N=36) played repeated trials of a newly developed air-traffic control simulation game called MAHVS (Multiple Autonomous Heterogenous Vehicle Simulation) with an artificially intelligent (AI) cooperative partner. Trust was measured by participant acceptance rates of AI recommendations. Independent variables included anthropomorphism (low/high) and transparency (on/off). A "disaster" trial was implemented halfway through each block of trials, in which task difficulty increased dramatically and the AI failed to help, after which an apology and explanation was offered by the agent. Results showed a significant effect of transparency on trustworthiness. Despite the disaster having a significant negative impact on trust, we found a significant interaction between transparency and the disaster, suggesting that transparency was most beneficial for trust repair following the disaster. Interestingly, despite competence being equal in all conditions, participants perceived the high-anthropomorphic AI to be less competent and less trustworthy than the low-anthropomorphic AI.