From low invasiveness to high control: how artificial intelligence allows to generate a large pool of standardized corpora at a lesser cost

Abstract

The use of corpora represents a widespread methodology in interpersonal perception and impression formation studies. Nonetheless, the development of a corpus using the traditional approach involves a procedure that is both time- and cost-intensive and might lead to methodological flaws (e.g., high invasiveness). This might in turn lower the internal and external validities of the studies. Drawing on the technological advances in artificial intelligence and machine learning, we propose an innovative approach based on deepfake technology to develop corpora while tackling the challenges of the traditional approach. This technology makes it possible to generate synthetic videos showing individuals doing things that they have never done. Through an automatized process, this approach allows to create a large scale corpus at a lesser cost and in a short time frame. This method is characterized by a low degree of invasiveness given that it requires minimal input from participants (i.e., a single image or a short video) to generate a synthetic video of a person. Furthermore, this method allows a high degree of control over the content of the videos. As a first step, a referent video is created in which an actor performs the desired behavior. Then, based on this referent video and participant input, the videos that will compose the corpus are generated by a specific class of machine learning algorithms such that either the facial features or the behavior exhibited in the referent video are transposed to the face or the body of another person. In the present paper, we apply deepfake technology to the field of social skills and more specifically to interpersonal perception and impression formation studies and provide technical information to researchers who are interested in developing a corpus using this innovative technology.