Mechanism behind preserved self-confidence after prior unlimited-time human-AI interaction

Establish whether initial unlimited-time human-AI interaction with the reinforcement learning Pac-Man agent enables users to understand the impact of their own contributions to the joint problem-solving process and to value themselves accordingly, thereby explaining the observed difference in self-confidence under subsequent time pressure.

Background

The study compares two sequences of mixed-initiative human-AI collaboration with an RL-based Pac-Man agent: one group first interacted without time pressure and then under time pressure, while the other experienced the reverse order. A significant decrease in self-confidence was observed only when time pressure came first, whereas starting without time pressure appeared to preserve self-confidence.

To explain this sequence effect, the authors conjecture that an initial, unhurried interaction helps participants appreciate the value and effect of their own contributions within the joint human-AI problem-solving process. Validating this mechanism would clarify why prepared interaction mitigates the adverse self-confidence impact of urgency and would inform onboarding and training strategies for agentic AI systems.