Risk-Free Study Designs Miss the Weak Trust–Behavior Link in Human–Automation Interaction: Evidence From Two Virtual Reality Experiments

Theories and practitioners alike assume that trust—a latent construct—causally predicts behavior. In human–automation and human–AI interaction, a system's characteristics (e.g., its reliability) are often thought to shape trust, which in turn supposedly predicts how much a human depends on the system. However, little evidence thus far supports this claim, particularly under risk, which theory treats as a precondition for a trust–dependence link. We addressed this gap in two laboratory experiments that used virtual reality to manipulate risk immersively, as part of the environment. In Experiment 1, we varied system reliability (low vs. high) and risk (absent vs. present) between subjects and tested whether several trust measures mediated the effect of reliability on dependence behavior. Trust mediated this effect only under risk, and only for a single-item measure. Thus, a standard no-risk design would have missed the link entirely. However, even the single-item trust measure explained little of the variance in behavior. In Experiment 2, conducted entirely under risk, we replicated the single-item trust mediation but found two further mediators: external regulation (the felt pressure to monitor the system) and one's perceived competence relative to the system. None of them, however, explained more than a small portion of variance within dependence behavior. Thus, even where trust predicted behavior, it did so weakly and non-uniquely. Subjective measures therefore cannot substitute for measuring behavior—whether or not trust theory is correct.