What Is Generative AI?

For most of its history, applied machine learning answered closed questions: is this email spam, which face is this. Generative AI answers an open one, producing an email, an image, a paragraph of code that did not exist a moment before. That shift is why a technology that once sat quietly inside recommendation feeds now writes, draws, and converses where anyone can see it. A tool that manufactures fluent, confident output is enormously useful and easy to over-trust, and the gap between plausible and correct is exactly where careful thinking earns its keep.

From discriminative to generative

Earlier lessons described models that learn patterns to assign a label, a discriminative task. Generative models learn patterns to produce a new example that fits them. A discriminative model looks at a photo and says "cat"; a generative model, given the word "cat," produces a photo of one that never existed. Both rest on the same foundation of learning statistical patterns from data. They simply point that machinery in opposite directions, one toward recognition, one toward construction.

What "generate" actually means

The word invites misunderstanding. A generative model does not invent from nothing, and it does not retrieve a stored answer. It samples from the statistical patterns it absorbed in training, assembling output piece by piece so each part is a likely continuation of what came before. The result is genuinely novel, not a file copied from the training set, yet entirely derived from that set's regularities. It is recombination at scale, closer to a remix than to imagination.

One family, many media

Text, images, audio, and code look like different problems but share one recipe: learn the patterns in a large collection of examples, then produce new samples that obey them. A language model predicts likely sequences of words; an image model predicts likely arrangements of pixels. The medium changes; the underlying move, model a distribution and draw from it, does not.

Ask an image generator for "a teapot shaped like a hedgehog." No such photo exists in its training data, yet it produces one. It has learned separately what teapots and hedgehogs tend to look like, and blends those patterns into a coherent new picture. Nothing was copied and nothing was imagined; the novelty comes from combining familiar regularities the data never joined.

A plain spam filter is not generative, however advanced. Show it a message and it returns one of two labels, spam or not, a decision from a fixed set. It never authors a message of its own. Sorting inputs into existing bins, however accurately, is discrimination, not generation. The line is not how clever the system is but whether its output is a choice among categories or a freshly built artefact.

Generative techniques existed in research labs for years, but they reached the public in a rush. The late-2022 release of ChatGPT put a conversational text generator in front of millions, and image generators reached wide audiences over the same period. What changed was less a single breakthrough than accessibility: models trained to produce text and images became usable through a simple box anyone could type into. The anchor is useful because it is verifiable and dated, but the durable lesson is the capability that became public, not any one product.

• "Generative AI creates from nothing." — It recombines patterns learned from training data; every output is derived from what it was shown.
• "The model copies pieces from its training set." — Typical outputs are new constructions, not retrieved files, though they mirror the data's regularities.
• "Generating content means the system is creative like a person." — It optimises for statistical plausibility, with no intent, taste, or understanding behind the output.
• "Generative and discriminative are rival technologies." — They are two uses of the same pattern-learning foundation, aimed at construction versus recognition.

You are shown a series of jobs, such as flagging toxic comments, drafting a product description, ranking search results, and composing a melody, and you decide which are discriminative and which are generative, defending each call.

Think Like a Maester: A generative model does not know things; it produces what is likely. Fluency is manufactured from patterns, so treat a polished output as a plausible draft to verify, never as a settled fact.

1. In your own words, how does a generative model differ from a discriminative one?
2. Why is it wrong to say a generative model either copies its output or invents it from nothing?
3. What do text, image, and audio generation share underneath?
4. Give an example of a task that is discriminative, not generative, and say why.
5. What actually changed for the public around late 2022?

Generative AI is the branch of machine learning that produces new content, text, images, audio, code, instead of only sorting inputs into categories. It builds on the pattern-learning foundation of earlier machine learning but aims it at construction rather than recognition, sampling from learned patterns to assemble outputs that are novel recombinations, not copies or human creativity. The public wave since late 2022 made this visible to everyone at once. The lasting skill is to see fluent output as a plausible draft shaped by training data, worth verifying, never mistaken for knowledge.

• What Is Machine Learning?
• How Large Language Models Work