What Is Generative AI? A Plain-Language Guide

The distinction that matters

For most of machine learning's commercial history, models sorted things. Is this email spam, is this transaction fraudulent, will this customer churn — each takes an input and returns a label or a number. Generative AI inverts that. Instead of judging content that already exists, it produces content that did not: a paragraph, an image, a function, a voice track. That single change in output type is why the technology arrived in products so suddenly, and why it fails in unfamiliar ways.

How it works, roughly

A generative model is trained to predict what comes next. A language model learns, over enormous quantities of text, to guess the next token given everything before it; run that guess repeatedly and you get sentences. An image model learns to reverse a noising process, starting from random static and refining it step by step until it resembles the description you gave. Neither is retrieving a stored answer. Both are producing a statistically plausible continuation, which is the source of both their flexibility and their unreliability.

What it is genuinely good at

The technology excels where a plausible draft is more valuable than a correct answer. Summarising a long document, rewriting text in a different register, translating, generating code for a well-trodden pattern, extracting structured fields from messy prose, producing a first version of anything a human will then edit. In each case the output is checked before it matters, and the model has converted a blank page into something to react to — which is usually the expensive part of the work.

Where it reliably disappoints

It disappoints when correctness is not verifiable at a glance. A model asked for a specific fact, a precise number, a legal citation, or a current price will produce something correctly shaped and possibly wrong, with no signal distinguishing the two. It does not know what it does not know. It is also weak at tasks requiring exact reasoning over many steps, and inconsistent across runs — ask twice, get two answers. None of that is a bug awaiting a patch; it follows from predicting plausible continuations rather than deriving results.

The practical consequence for a product

The useful question is never whether the model is impressive. It is what happens when the output is wrong. A summary a human reads next to the source is safe, because the error is visible and cheap. The same model deciding a refund, sending an email unattended, or writing to a database is not, because the error escapes before anyone sees it. Most disappointing AI features are not badly built; they are correctly built around a task that could not tolerate being wrong.

Cost and latency are design constraints

Generative models bill per token and respond in seconds, not milliseconds. That rules out some patterns entirely — you cannot call one on every row of a large table, or inside a request that must return instantly. It also means the shape of your prompt has a running cost, so context that is convenient to include is not free. Teams routinely discover this after launch, when a feature that tested fine on a hundred users becomes the largest line on the infrastructure bill.

How to evaluate it honestly

Build a small fixed set of representative inputs with known-good outputs before you build the feature, and score against it whenever you change the prompt or the model. Without that, "it seems better" is the only available measure, and it is not a measure. This is unglamorous and it is the difference between a feature you can improve deliberately and one you tune by vibes.

The short version

Generative AI turns a blank page into a draft, cheaply and at any hour. It does not turn an uncertain question into a reliable answer. Products that respect that distinction tend to work; products that blur it tend to ship, impress in a demo, and quietly get switched off.