Getting Emotional Enough : Analyzing Emotional Diversity in Deepfake Avatars

Abstract

When using deepfake technology to represent users, there is a need to convey a reasonable range of emotions to be able to portray different circumstances ranging from positive to negative experiences (e.g., personal struggles). Because it is not known how well deepfake avatars embody emotional diversity, we investigated this aspect among 202 deepfake avatars. Our findings suggest an overall positivity bias in deepfake avatars’ emotions. We also found significant gender differences in several emotional expressions, with, male deepfakes scoring higher in “smile” and “calm” emotions, and female deepfake avatars scoring higher in “surprised”, “fear”, and “happy” emotions. In terms of ethnicity, European and Hispanic deepfake avatars demonstrate the broadest range of “smile”, “happy”, and “calm” compared to other ethnic groups. Age had no notable bias. No emotion score was normally distributed, suggesting that the range of emotional representativeness among the tested deepfake avatars is skewed. We outline the implications for academics and professionals regarding future development and responsible deployment of deepfake avatars.