We investigate the usability of human-like agent-based interfaces. In an experiment we manipulate the capabilities and the “human-likeness” of a travel advisory agent. We show that users of the more human-like agent form an anthropomorphic use image of the system: they act as if the system is human, and try to exploit typical human-like capabilities. Unfortunately, this severely reduces the usability of the agent that looks human but lacks human-like capabilities (overestimation effect). We also show that the use image users form of agent-based systems is inherently integrated (as opposed to the compositional use image they form of conventional GUIs): cues provided by the system do not instill user responses in a one-to-one manner, but are instead integrated into a single use image. Consequently, users try to exploit capabilities that were not signaled by the system to begin with, thereby further exacerbating the overestimation effect.
Knijnenburg, Bart and Willemsen, Martijn C., "Inferring Capabilities of Intelligent Agents" (2014). SIGHCI 2014 Proceedings. 9.