AI-generated images of familiar faces are indistinguishable from real photographs.

Abstract

Human users are now able to generate synthetic face images with artificial intelligence (AI) tools. Although indistinguishable from real photographs, these images have tended to feature fictional identities that do not exist in the real world. As a result, their use in applied contexts, including the spread of fake information, is similarly limited. Here, we investigated a new method for generating face images (via ChatGPT plus DALL·E) and its application to both fictional and real (in this case, celebrity) identities. Our results demonstrated that generated images of both fictional (Experiment 1) and celebrity identities (Experiment 2) could not be distinguished from real photographs. Further, providing additional real photographs for comparison during the task resulted in limited gains (Experiments 3 and 4). Finally, prior familiarity with celebrity faces produced only modest performance improvements. Therefore, new methods of detection should be explored as a matter of urgency since the latest 'off the shelf' AI tools can now generate face images of real people that are essentially undetectable as synthetic to most human observers.