Two neural networks trained against each other, one producing synthetic data and the other judging whether it is real.
In a Generative Adversarial Network (GAN), one network forges examples and a second network tries to catch the forgeries, and each improves by competing with the other. This adversarial loop has been widely used to generate realistic synthetic data, for example to train autonomous systems or to augment datasets in privacy-constrained environments.
See also Synthetic Data, Large Language Model, Fine-Tuning and Model Inference.