A GAN is a machine learning model with two competing neural networks: a generator creating new data and a discriminator classifying it as real or fake. They improve through adversarial training, producing highly realistic synthetic data.