Abstract
Dynamic facial expressions can be created by blending neutral and expressive photographs together. These dynamic morphs are popular in cognitive neuroscience research as they offer increased experimental control compared to video recordings, which can vary in the duration and temporal profile of each expression. However, emerging research suggests that morphed motion may not capture the temporal dynamics of real facial motion. Deepfake technology offers a potential solution this issue by using deep learning algorithms to transpose one face onto a video of another, while preserving the facial motion. This allows for the creation of facial expressions that exhibit multiple identities with the same spatiotemporal characteristics. Real videos were compared with static photographs, dynamic morphs, and deepfakes that were all created from the same set of video-recorded emotional expressions. We used electroencephalography (EEG) to measure neural and behavioural responses to emotional expressions exhibited in these four presentation types. Morphed expressions of emotion (happiness, anger, fear, and sadness) were perceived as less strong compared to video recordings, static photographs, and deepfakes. Morphed happy expressions were also rated as less genuine than those in video, static, and deepfake formats. Strength and genuineness ratings did not differ between videos and deepfakes. Visual evoked potentials revealed that compared to photographs and videos, dynamic morphs produced decreased amplitudes in the late positive potential (LPP) component, which is associated with recognition and evaluation of emotion intensity. These results suggest that linear morphs may not be a suitable replacement for video recordings. Deepfake technology may offer researchers the ability to create realistic dynamic stimuli that can be manipulated for specific experiments