Using Deepfakes for Experiments in the Social Sciences: Studies on Discrimination

Since the end of 2017, the creation and distribution of deepfakes have increased sharply. Deepfakes provide the ability to swap one person’s face onto another in a picture or a video based on artificial intelligence applying deep learning techniques. The specific algorithm, which creates these fake videos, learns and improves by constantly mimicking gestures and facial expressions. Since this technology develops continuously, the results are also improving steadily, making deepfake videos indistinguishable in quality from the original videos.

Besides the use of this technology for entertainment and manipulation purposes, we see great potential in using deepfakes for substantive research – especially as an approach for experimental manipulation with a high degree of control in the social sciences. Thereby, we argue that deepfakes can be a valuable tool for conducting social science experiments. In several projects, we aim to examine discrimination in occupational contexts based on gender, physical attractiveness or ethnicity. This technology allows us to systematically vary visual and auditory stimulus material in experiments while holding everything else constant. For example, we can manipulate a person’s face while keeping all other video elements identical, such as the audio recording and its speed, the background, clothing, and hairstyles. Consequently, deepfakes provide a high degree of experimental control and thus appear to be a promising method for identifying causal effects in experiments by systematically varying only one factor at a time.

Project team: Tobias Wolbring, Juliane Kühn, Andreas Eberl

Funding: Intern, Emerging Talents Initiative

Publications:

Eberl, A., Kühn, J., & Wolbring, T. (2022). Using deepfakes for experiments in the social sciences–A pilot study. Frontiers in Sociology, 7, 907199. https://dx.doi.org/10.3389/fsoc.2022.907199
DOI: 10.3389/fsoc.2022.907199