What is a Generative AI?

Earlier applied AI was mostly discriminative: given an input, predict a label (spam or not spam, fraud or not fraud). Generative AI instead learns the underlying distribution of its training data well enough to produce new, original outputs that plausibly belong to that distribution — a language model writing a paragraph it has never seen before, a diffusion model producing an image from a text description, a code model writing a function it wasn't trained verbatim on. The large language models behind modern chatbots and AI agents are generative models applied to text and, increasingly, structured actions. What makes generative AI both powerful and risky is the same property: it produces plausible-sounding output even when it's factually wrong ("hallucination"), because it's optimizing for statistically likely continuations, not verified truth. This is why generative systems used for anything consequential are typically paired with grounding techniques — retrieval augmentation, tool use to check facts, human review — rather than trusted to generate correct answers unaided. The category now spans far beyond chatbots: generative models draft marketing copy, synthesize code, produce images, and drive the reasoning loop inside autonomous agents.