Room 4 · The Awakening · 2014

Generative Adversarial Networks

Generative Adversarial Networks (Ian Goodfellow and colleagues, 2014): Two networks that compete and improve each other.

In a GAN, a generator tries to create fake images that look real and a discriminator tries to tell them from the real ones. By competing, both improve, and the generator ends up producing ever more convincing images.

In June 2014 Ian Goodfellow and colleagues at the University of Montreal, in the group led by Yoshua Bengio, posted "Generative Adversarial Nets." The paper was presented at the NIPS conference that December and introduced one of the most influential ideas in generative AI before diffusion models.

A GAN has two networks with opposite goals. The generator starts from random noise and tries to produce data, such as images, that look real. The discriminator receives a mix of real images and the generator's images and tries to tell which is which. Each network is trained against the other. When the discriminator gets better at spotting fakes, the generator has to improve to fool it, and when the generator improves, the discriminator has to sharpen its judgment. Ideally, the contest ends when the generator's images are indistinguishable from real ones.

The first results were small, blurry images of digits and faces. Over the following years, improvements in the architecture and the training made progress fast. By 2019 the StyleGAN networks from Nvidia produced photographic faces of people who do not exist, convincing enough that a website showing a new one on every visit went viral.

GANs were also used to turn sketches into photos, to change the season in a landscape, to restore old images and, controversially, to create deepfakes: fake videos and images of real people. That last use made GANs part of the public debate about synthetic media and consent, which you can follow in our Culture and Safety sections.

For image generation, diffusion models later took the lead, as you will see in Room 5. The adversarial idea remains a classic.

Ask your AI: explain a GAN as a contest between a counterfeiter and a detective, then explain where the analogy fails.

Did you know? According to Goodfellow, the idea came to him in a bar, arguing with friends, and he tested the code that same night.

Source: arxiv.org/abs/1406.2661

See it in the museum →

← word2vec: Meaning as Geometry · DeepMind Learns to Play Atari →

All of Room 4: The Awakening

Cite this: Nazareno, I. (2026). Generative Adversarial Networks (2014). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/generative-adversarial-networks-2014/