Registered Report: Does the valence-dominance model apply to deepfake faces? Validation and its impact on deepfake detection

Abstract

Artificial Intelligence has given rise to increasingly more realistic fake contents such as deepfake faces, hardly distinguishable from real faces. However, evidence has shown that social evaluations of real faces are also able to impact the detection of deepfake faces. The present study aimed at applying for the first time the valence-dominance model for face evaluation on deepfake faces. Moreover, the impact of its thirteen attributes on deepfake detection performances was investigated. Specifically, 311 participants categorised 120 faces (50% real, 50% deepfake created ad hoc by Diffusion Model technology) and rated faces in regards to one of the attributes. Principal Component Analysis was applied in order to validate the valence-dominance model on deepfake faces, and a Brunswikian lens model was applied to understand how each attribute contributes to a face being judged as deepfake. Results replicated the valence-dominance model on real faces as well as on faces categorized by participants as deepfake, but not on actual deepfake faces. Finally, the indirect effects of the Brunswikian model revealed caringness and intelligence to impair detection of deepfake stimuli, whilst weirdness was revealed to be a positive indirect predictor, being positively associated with higher detection accuracy of deepfake faces. Our results shed, for the first time, light into the differential cognitive processing of deepfake faces and their social evaluation whilst also providing novel insight into underlying mechanisms associated with deepfake detection. Interdisciplinary implications regarding the theoretical and applied contribution are discussed.