A generative adversarial network, or GAN, trains two neural networks against each other. A generator tries to create realistic fake data, such as images, while a discriminator tries to tell the fakes from real examples. As they compete, the generator gets better at producing convincing output until its creations are hard to distinguish from the real thing.
GANs were introduced by Ian Goodfellow and co-authors, including Yoshua Bengio, in the 2014 paper “Generative Adversarial Networks.” The paper shows that the system has a unique optimal solution where the generator perfectly reproduces the real data distribution and the discriminator can do no better than guessing, and that the whole thing can be trained with backpropagation without Markov chains.
GANs drove a wave of progress in synthetic media, including photorealistic faces, image enhancement, and style transfer, and they introduced the public to both the creative potential and the deepfake risks of generative AI.
Why business readers should care: GANs were an early demonstration that AI could create, not just classify. They underpin tools for image generation, data augmentation, and product visualization, while also raising the authenticity and trust concerns that now shape AI policy.