Study: Users trust AI agents per task, not per system — leading to 'delegation regret'

Researchers at a university conducted a controlled study where 20 students used OpenClaw, a general-purpose AI agent, to complete five common daily tasks varying in privacy, stakes, and reversibility. They measured trust, perceived control, transparency, supervision burden, and approval preference on 5-point Likert scales, plus collected free-text reflections.

Three key findings emerged. First, participants calibrated trust per task rather than per agent: they granted wide autonomy for advisory and low-stakes tasks but demanded confirmation for irreversible, externally visible actions. Second, irreversibility combined with external visibility — not stakes alone — drove trust withdrawal: the moderate-stakes email task produced the sharpest trust drop (M=3.10) and highest approval demand (M=4.65), while a high-stakes but verifiable task did not trigger the same response. Third, 'delegation regret' appeared consistently when the agent executed actions without preview, even when the output was rated as successful.