Training super-recognisers’ detection and discrimination of computer-generated faces

Generative Adversarial Networks (GANs) can create realistic synthetic faces, which have the potential to be used for nefarious purposes. The synthetic faces produced by GANs are difficult to detect and are often judged to be more realistic than real faces. Training programmes have been developed to improve human GAN detection accuracy, with mixed results. Here we investigate GAN face detection and discrimination in a sample of super-recognisers (who have exceptional face recognition skills), and typical-ability participants. We also devised a training procedure which focused on highlighting rendering artifacts. In two different experimental designs, we found that super-recognisers were better at detecting GANs than controls, but their performance did not differ from chance, whereas control participants’ performance was significantly below chance. Trained super-recognisers and controls had significantly better performance than those without training, and the magnitude of the training effect was similar in both groups. With training, super-recognisers performed above chance, whereas controls improved to the extent that they performed at chance. Our findings suggest that super-recognisers are detecting GANs using cues that are unrelated to rendering artifacts, and their performance can be enhanced through a short training procedure. These results have implications for synthetic image detection in real-world scenarios.