Behavioral and Mechanical Responses to Hyper-realistic Digital Humans Faces (Still Life)

Abstract

My project investigated whether differences in the perceived realism level of faces would cause differences in trait evaluations. I conducted one survey questionnaire study (N = 299) and one eye-tracking study (N = 41). Subjects rated various traits under different occupational contexts in a series of faces differing in realism level. Our survey findings show that detected non-humanness did not disadvantage less realistic faces compared to human faces during trustworthiness, officialness, and competency trait ratings. The perceived likability of a face was not affected by the face's realism level; thus, digital human faces were rated significantly more likable than human faces. We did not find evidence for systematic changes across humanity across most contexts. Subjects across all occupational contexts tended to rate digital human faces and human faces similarly. Interestingly, we found the participants only tended to care about the humanity differences across the different occupational risks during the referenceless agent preference rankings, but not in the referenced trait ratings. Eye-tracking further provided mechanistic insights: the detection of non-humanness happened as early as during the second fixation to the face. Fixations directed to non-human faces were typically prolonged compared to human faces. These differences, however, did not reside in specific regions of the face. Together, this project showed that human observers have a default tendency to detect non-humanness during onsets of less realistic faces, which plays an important role in the formation of trait judgments across different occupational contexts.