Effects of Referring to Robot vs. User Needs in Self-Explanations of Undesirable Robot Behavior

Abstract

Autonomous or lively social robots will often exhibit behavior that is surprising to users and calls for explanation. However, it is not clear how such a robot behavior should be explained best. Our previous work showed that different types of a robot's self-explanations, citing its actions, intentions, or needs -alone or in causal relationshave different effects on users [19]. Further analysis of the data from the cited study implies that explanations in terms of robot needs (e.g. for energy or social contact) did not adequately justify the robot's behavior. In this paper we study the effects of a robot citing the user's needs to explain its behavior. Our study is based on the assumption that users may feel more connected to a robot that aims to recognize and incorporate the users' needs in its decision-making, even when the resulting behavior turns out to be undesirable. Results show that explaining robot behavior with user needs generally did neither lead to higher gains in understanding or desirability of the behaviors, nor did it help to justify them better than explaining it with robot needs. Further, a robot referring to user needs was not perceived as more likable, trustworthy or mindful, nor were users' contact intentions increased. However, an in-depth analysis showed different effects of explanations for different behaviors. We discuss these differences in order to clarify which factors should inform content and form of a robot's behavioral self-explanations. CCS CONCEPTS• Human-centered computing → Empirical studies in HCI; User studies; Scenario-based design; Empirical studies in interaction design.

Explaining Before or After Acting? How the Timing of Self-Explanations Affects User Perception of Robot Behavior

“…Further exploratory analyses of participants' ratings of the behavior explanations revealed no significant differences between explanation timing groups, supporting Q3c-e, while negating Q3a&b. Still, and in line with previous results [18,19], the differences between the behaviors have a statistically significant impact on how the explanations are rated, suggesting that the effect of explanations should always be considered in the context of the specific explanandum.…”

“…We chose to investigate the effect of explanation timing with a subset of our previously evaluated video stimuli showing the Pepper robot. We included only behaviors that originate from internal robot needs and, in previous work, were rated as most surprising and undesirable [18,19]. Namely, the following three scenarios were selected (original video number in parenthesis) 3 -The robot playfully blocks the user's way (need for social contact) [8] -The robot drives into the picture and blocks the user's view to the TV (need for social contact) [11] -The robot begins to sing and dance, while the user is asleep on the couch (need for entertainment) [12] Unlike in our previous studies, the robot did not explain its behavior upon request, but rather explained it proactively.…”

“…Positive effects of explanations in human-AI interaction have been reported in numerous studies. Explanations can increase understandability of tasks, help to build adequate trust, and increase desirability of undesired behaviors [11,18,19]. While the question of what an explanation should entail has been tackled in multiple fields, the question of when to provide an explainee with an explanation has received relatively little attention in human-robot interaction so far.…”

“…In order to justify an agent's choice, for instance, the information can either be presented before the agent acts, in a preventative manner, or afterwards (accusatorily). Still, a social robot's beliefs and goals are often explained after the behavior, which has been found to increase the perceived desirability of undesirable social robot behaviors [18,19]. We generally adopt the view of explaining as a dialogical process [21] that responds to an explainee's need for explanation.…”