Calibrating trust to integrate intelligent agents into human teams

Abstract

Complex team tasks, such as joint operation planning, require that information and resources from distributed sources be exchanged and fused because no one individual or service has the collective expertise, information, or resources required. We are at the beginning of a major research program aimed at effectively incorporating intelligent agents into human teams. Our initial experiments use a low fidelity simulation of a target identification task, TANDEM, to investigate human-agent interaction for individual operators and human teams. In the TANDEM simulation subjects identify and choose an appropriate response to a series of air, surface, or submarine targets. Subjects perform a sequence of time critical information gathering and communications tasks in order to decide whether to shoot or clear each target. The data necessary to make a correct decision is distributed among human and in our experiments software agents. Subjects must communicate, coordinate, and negotiate with one another and their agents in order to amass the information needed to classify and dispose of each target. The most transparent presentation led to the best performance and greatest reliance on the simulated agent.