When Transparent does not Mean Explainable

Abstract

Based on findings from interactional linguistics, I argue thattransparency is not desirable in all cases, especially not in socialhuman-robot interaction. Three reasons for limited use oftransparency are discussed in more detail: 1) that social humanrobotinteraction always relies on some kind of illusion, whichmay be destroyed if people understand more about the robot’s realcapabilities; 2) that human interaction partners make use ofinference-rich categories in order to inform each other about theircapabilities, whereas these inferences are not applicable to robots;and 3) that in human interaction, people display only informationabout their highest capabilities, so that if robots display low-levelcapabilities, people will understand them as very basic. I thereforesuggest not to aim for transparency or explainability, but to focuson the signaling of affordances instead