What is Generative AI?

Producing new content, not picking a label

For most of its history, machine learning was about choosing: is this email spam, which of ten categories is this image, what number comes next on this chart. The answer always came from a list fixed in advance.

Generative models produce the content itself — a paragraph, a picture, a block of code, a voice. They learn what real examples look like closely enough to produce new ones that fit, building them one piece at a time: the next word, the next patch of an image.

That shift is what made AI usable by anyone who can type a request. It also changes what goes wrong. A classifier that is wrong picks the wrong label from a short list; a generator that is wrong produces something fluent, plausible and new, which is much harder to spot. The page on hallucination is about exactly that.