It's best known from generative adversarial networks, where the generator learns to fool the discriminator with a handful of convincing outputs rather than exploring the whole distribution. The same term gets used loosely for language and image models that default to repetitive, samey outputs across different prompts.