GAN Generative Adversarial Network
Two neural networks compete: Generator creates fake data, Discriminator tries to spot the fake. Both get better until the generator's output is indistinguishable from real.
Ian Goodfellow, 2014. Drove the 2014–2020 wave of AI-generated images, "deepfakes", StyleGAN (NVIDIA, 2018, photorealistic faces). Notoriously hard to train stably — mode collapse, training instability. Replaced 2021–2024 by diffusion models (Stable Diffusion, DALL·E 2+) which give higher quality with more stable training. Lives on in special cases (super-resolution, style transfer).