Generative AI refers to machine learning systems whose primary output is new content — text, images, audio, video, code, or structured data — synthesized by learning statistical patterns from large training corpora. Unlike discriminative models (which classify or rank existing inputs), generative models build an internal representation of the data distribution and then sample from it to produce novel artifacts. The main architectures driving the current wave are autoregressive transformers (for text and code), diffusion models (for images and video), and flow-matching models; older approaches include variational autoencoders and GANs.

Think of it like the difference between a film critic and a director. A critic (discriminative model) watches a movie and labels it "good" or "bad." A director (generative model) has absorbed thousands of films and can produce something new that looks and feels like a real movie — it generates rather than classifies.

A minimal example: asking an LLM to draft code, then piping that to an image model to create a diagram, illustrates how generative AI components chain together in a pipeline:

import anthropic, requests
client = anthropic.Anthropic()
code = client.messages.create(
    model="claude-opus-4-5", max_tokens=512,
    messages=[{"role":"user","content":"Write a Python quicksort."}]
).content[0].text
# The text itself is generative output — novel tokens sampled from the model

Generative AI encompasses LLM (text generation), Diffusion (image/video synthesis), GAN (adversarial generation), Multimodal systems that generate across modalities, and Foundation Model (the large pretrained base these systems build on). Temperature is a key control knob governing how random or conservative the outputs are.

You will encounter the term in product marketing, policy discussions, and technical literature whenever the conversation concerns AI systems that produce rather than merely analyze content — which, since 2022, has become the dominant context for mainstream AI discussion.