Expressions of robot confidence and their effect on trust and acceptance of recommendations

Verbal and non-verbal cues between humans are important for trust formation and relationship development. This study compared a humanoid robot's probabilistic verbal expressions of confidence (e.g., "I am 95% confident") versus non-numeric descriptive verbal expressions of confidence (e.g., "I am very confident"), while also comparing pre-programmed facial expressions of confidence versus teleoperated motion capture expressions of confidence (e.g., no two expressions are identical). Participants (n=26) played the 'shell game' with a Hanson RoboKind robot 'Milo', who made recommendations as to the correct answer. Multilevel Bayesian regression found very strong evidence that confidence had a significant effect on recommendation acceptance, and that participants trusted the robot irrespective of whether there were gains in task performance or not. Most importantly, we found that participants were more likely to trust the robot when the robot used probabilistic verbal expressions of confidence (e.g., "95%") as compared to descriptive expressions of confidence (e.g., "very").