Testing human ability to detect ‘deepfake’ images of human faces

Abstract ‘Deepfakes’ are computationally created entities that falsely represent reality. They can take image, video, and audio modalities, and pose a threat to many areas of systems and societies, comprising a topic of interest to various aspects of cybersecurity and cybersafety. In 2020, a workshop consulting AI experts from academia, policing, government, the private sector, and state security agencies ranked deepfakes as the most serious AI threat. These experts noted that since fake material can propagate through many uncontrolled routes, changes in citizen behaviour may be the only effective defence. This study aims to assess human ability to identify image deepfakes of human faces (these being uncurated output from the StyleGAN2 algorithm as trained on the FFHQ dataset) from a pool of non-deepfake images (these being random selection of images from the FFHQ dataset), and to assess the effectiveness of some simple interventions intended to improve detection accuracy. Using an online survey, participants (N = 280) were randomly allocated to one of four groups: a control group, and three assistance interventions. Each participant was shown a sequence of 20 images randomly selected from a pool of 50 deepfake images of human faces and 50 images of real human faces. Participants were asked whether each image was AI-generated or not, to report their confidence, and to describe the reasoning behind each response. Overall detection accuracy was only just above chance and none of the interventions significantly improved this. Of equal concern was the fact that participants’ confidence in their answers was high and unrelated to accuracy. Assessing the results on a per-image basis reveals that participants consistently found certain images easy to label correctly and certain images difficult, but reported similarly high confidence regardless of the image. Thus, although participant accuracy was 62% overall, this accuracy across images ranged quite evenly between 85 and 30%, with an accuracy of below 50% for one in every five images. We interpret the findings as suggesting that there is a need for an urgent call to action to address this threat.