Term

Generative Adversarial Network (GAN)

A Generative Adversarial Network is a generative model built from two networks in opposition. A generator produces synthetic samples (an image, an audio clip, a tabular row) and a discriminator judges real versus fake. Both improve through the adversarial training loop: the generator learns to fool, the discriminator learns to catch. Introduced by Goodfellow and co-authors in 2014. The architecture dominated image generation from roughly 2017 to 2022 (StyleGAN, BigGAN) before diffusion models displaced it in most consumer-grade work.
Reviewed by 7wData

Why it matters

Even with diffusion’s rise, GANs are still in use where speed or tight control beats peak fidelity: real-time generation, conditional synthesis, several biomedical and industrial pipelines. The failure modes (mode collapse, training instability) are pedagogically foundational. For AI governance, GANs are the textbook “model that can fabricate synthetic data”, which is why they sit at the centre of the early deepfake, synthetic-data privacy, and adversarial-input training debates. Most “deepfake” coverage from 2018 to 2022 was GAN-driven; from 2023 the conversation shifted to diffusion plus LLM-controlled generation, but the policy vocabulary was set in the GAN era.

Where you’ll encounter it

Three concrete contexts. A model-risk team asks about deepfake exposure, GAN history is the load-bearing context for why the controls exist. A vendor pitches “synthetic training data via GANs”, sometimes legitimate, sometimes diffusion in a GAN-shaped costume because the buyer expects the older word. A research review benchmarks generative-quality metrics (FID, IS) developed in the GAN era and still quoted on diffusion outputs. The pitfall I see most often: people conflate GANs with diffusion because both are “generative” and produce images. Different architectures, different failure surfaces, and treating them as interchangeable is how compliance reviews answer the wrong question.

Part of the 7wData AI Glossary. Tracking how concepts like this move in the expert conversation: daily signals at ins7ghts.com.