An interactive web app that asks visitors to act as the discriminator for a generative adversarial network (GAN). Visitors distinguish between real photos and AI-generated faces, and their feedback improves the model from noise to near-photorealistic output. Scoring tracks accuracy, sensitivity, and specificity, showing how those three measures are used in combination to understand AI performance.

Face outputs are from three pre-trained models used to visualize GAN output across epochs: a custom TensorFlow GAN trained on LFW, NVIDIA's StyleGAN and StyleGAN2, trained on CelebA and FFHQ.