Generative Adversarial Network (GAN)
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A deep learning framework in which two neural networks compete: one generates synthetic data, the other judges it, until the fakes pass for real.
What is a Generative Adversarial Network (GAN)?
A generative adversarial network (GAN) is a deep learning model built from two neural networks locked in competition: a generator that produces synthetic data and a discriminator that tries to tell fake from real.
Ian Goodfellow and colleagues introduced GANs in 2014. The design’s insight is that competition itself is a training signal: the generator improves precisely because the discriminator keeps catching it.
How a GAN Works
The generator starts from random noise and outputs a candidate sample, at first meaningless static. The discriminator receives a mix of real training examples and generated fakes, and labels each one real or fake.
Both networks learn from the verdicts. The generator adjusts to fool the discriminator; the discriminator adjusts to catch the improved fakes. Training succeeds when the discriminator does no better than a coin flip, because at that point generated samples are statistically indistinguishable from real ones.
Applications of GANs
- Image synthesis: photorealistic faces, scenes, and products that never existed, used in film, gaming, and design.
- Data augmentation: synthetic training examples for domains like medical imaging, where real labeled data is scarce or privacy-restricted.
- Deepfakes: swapped faces and cloned voices in video and audio, the most contested use of the architecture.
Example of a GAN
Train a GAN on a large collection of photographed human faces. Early in training the generator emits noise and the discriminator rejects everything. Round by round, the generator learns the geometry of faces: first blobs with eyes, then coherent features, eventually portraits with pores and stray hairs.
The finished generator produces a photorealistic portrait of a person who has never existed. The same capability powers deepfake impersonation, which is why the EU AI Act requires deepfakes to be disclosed as artificially generated.
Related AI terms: Generative AI · Variational Autoencoder · Synthetic Data · Data Augmentation
Did you like the Generative Adversarial Network (GAN) gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About