A GAN (Generative Adversarial Network) pits two networks against each other: a generator tries to produce realistic fake data, and a discriminator tries to tell real from fake. They train adversarially — the generator improves until its output fools the discriminator — yielding a generator that produces convincing samples.
Worked example: to generate faces, the generator makes images from noise while the discriminator scores them against real photos; the tug-of-war pushes the generator toward photorealistic faces. Gotcha: GAN training is notoriously unstable — mode collapse (the generator produces only a few outputs), non-convergence, and hyperparameter sensitivity — which is a big reason diffusion models, which train more stably, have largely overtaken GANs for image generation.