Trust calibration is critical for productive human-AI collaboration in production management, yet the impact of brief explanations on trust and reliance for black-box schedulers remains unclear. We conducted a between-subjects online experiment in flexible job shop scheduling (FJSS; N = 253 professionals and graduate engineers), comparing a deep reinforcement learning (DRL) scheduler augmented with natural-language rationales to a transparent first-in-first-out (FIFO) heuristic while holding recommendations identical to isolate explanation effects. In a five-step, path-dependent scheduling task, participants relied more on DRL with rationales, but attitudinal trust did not differ. Instead, trust increased through mediation of perceived ability. This indirect trust effect was smaller on harder tasks and larger among domain experts, whereas reliance was not moderated. Practically, short rationales function primarily as ability cues that raise adoption without necessarily deepening understanding. To maintain calibrated trust, these cues should be complemented with audience- and task-specific guidance on uncertainty and limitations.
Saßmannshausen et al. (Wed,) studied this question.