Guilty until proven innocent: Experiencing errors in training promotes relationship equity and trust calibration.

Area of Science:

• Human-computer interaction
• Artificial intelligence
• Cognitive psychology

Background:

• Trust is crucial for effective human-autonomous system collaboration.
• Understanding how system performance influences trust over time is essential for design.
• Previous research has not fully explored the impact of imperfect training on trust dynamics.

Purpose of the Study:

• To investigate the longitudinal development of trust in autonomous systems.
• To examine how an autonomous teammate's performance (perfect vs. imperfect) affects user trust, monitoring behavior, and task performance.
• To explore the implications of these findings for training autonomous systems.

Main Methods:

• Participants were assigned to work with either a perfect or imperfect autonomous teammate in a missile defense simulation.
• The study involved a training phase with varied teammate performance and a testing phase with consistent errors.
• Data collected included subjective trust ratings and objective performance metrics.

Main Results:

• Training with an imperfect autonomous teammate led to decreased trust and increased user monitoring.
• Participants who trained with an imperfect teammate demonstrated superior error prevention rates during the testing phase.
• This suggests that initial exposure to imperfection can enhance future performance and adaptive trust.

Conclusions:

• An imperfect autonomous teammate during training can foster greater vigilance and ultimately improve performance.
• Findings support a new theory of relationship equity in human-AI interactions.
• Implications for designing more effective training protocols for autonomous systems are discussed.