Picture a counterfeiter and a bank inspector. The counterfeiter (the “generator”) prints fake banknotes, and the inspector (the “discriminator”) learns to tell them from real ones. Each time the inspector catches a fake, the counterfeiter improves. Over many rounds the fakes become hard to distinguish from the real thing. GANs learned to produce realistic faces, artwork and photo edits this way.
GANs were introduced by Ian Goodfellow and colleagues in 2014 and drove much of the early progress in AI image generation. They can be tricky to train, because the two networks have to stay in balance. For many image tasks, diffusion models have since become the more common choice, though GANs are still used where speed matters.
Why it matters
GANs are a key chapter in the history of generative AI, and the adversarial idea, training one system against another, still shows up across machine learning.
No published Chain of Thought episode discusses GANs directly yet, so this entry relies on the original paper.
Related terms
A neural network is the building block for both halves of a GAN. A diffusion model is the main newer alternative.
Sources
- Goodfellow et al.: Generative Adversarial Networks — The 2014 paper proposing a framework in which a generative model and a discriminative model are trained simultaneously in an adversarial game; checked October 24, 2026.