Slang & Culture
Mode Collapse
When an AI model gets "stuck" producing only a narrow slice of the outputs it should be capable of — like an image generator asked for "a person" that keeps returning nearly identical faces.
Origin · no single documented coiner
The underlying failure was first described, though not by this name, in Ian Goodfellow and colleagues' original 2014 paper introducing Generative Adversarial Networks, which called it the "Helvetica scenario." The two-word term crystallized shortly after in Tim Salimans, Ian Goodfellow, and colleagues' 2016 paper "Improved Techniques for Training GANs" — no single person is credited with coining the exact phrase.