In simple words: A GAN learns to generate data through two competing networks.
Two networks
A Generative Adversarial Network (GAN) trains a generator to create synthetic examples and a discriminator to distinguish real examples from generated ones. They improve through this competing training process.
Possible uses and limits
GANs can create images or help with image restoration and research data augmentation. Synthetic images need quality checks. They can also be used to make deceptive content, so provenance and responsible use matter.
See the flow
generator makes sample ↔ discriminator checks sample
Simple example
A generator creates a fake handwritten digit. A discriminator tries to tell whether it came from the real digit dataset.
Remember: A GAN learns to generate data through two competing networks.