What is Generative AI? — one analogy

1The analogy

Walk into a flower shop and ask for a small bouquet, mostly yellow, small blooms only, and the owner doesn't hand you something already sitting finished on the shelf. They stand at the bucket of loose stems and build a new bouquet right there, adding one flower at a time. The bouquet they tie today is never quite the same as one made from the same request yesterday.

That's exactly what generative AI does. It isn't fetching an answer that was already made somewhere and stored — it takes the request and builds a new result, piece by piece, on the spot.

The hands move so surely because thousands of bouquets have already passed through them. Which leaf suits which flower is second nature by now, so even a request never heard before still puts the next stem within reach.

2In detail

Picking is different from building

What AI does splits broadly into two kinds. One is judging which category something belongs to — looking at a photo and answering whether it's a cat, or reading an email and sorting spam from not. The answer is always one of a fixed, known set.

The other is inventing something that didn't exist. There's no fixed list of answers to choose from. The sentences or pictures that could come out are effectively endless, and the job is to build one new one. That's as different from picking as choosing a bouquet off a shelf is from tying loose stems into one.

Judging what counts as success changes too. A category judgment can be scored right or wrong; an invented result has no answer key. All you can ask is whether it's plausible, and whether it fits the request.

Building it piece by piece

Generative AI doesn't produce a finished result in a single burst. Writing text, it picks the single next piece, attaches it, looks at the whole thing again, and picks the next piece after that — the same way a florist sets one stem, steps back to look at the whole bouquet, then reaches for the next.

Pictures work the same way. Instead of one finished stroke, the picture passes through dozens of steps, sharpening a little more out of a hazy starting state each time. Text streaming across your screen, or a picture gradually coming into focus, is that process actually happening in front of you.

Every piece gets picked with everything built so far in view. Set a red flower down early and it shapes every choice after it — which is exactly why one wrong pick early on can drag every later choice off course.

A sense for what fits, built from countless examples

Nobody wrote down a rule for what word follows what, or what color belongs where. That sense comes from combing through enormous amounts of text and pictures and picking up what patterns kept showing up together — what word tends to follow another, what color a sky tends to be near the top.

That sense is different from memorizing whole sentences or pictures outright. Memorizing would leave it helpless in front of anything new. Because what it learned is how pieces fit together, it can still produce something for a combination it's never seen before.

That's also exactly where the limits come from. A flower that was never in the bucket can't end up in the bouquet — something barely represented in training comes out looking rough. That's why hands and lettering, both common subjects with unusually strict rules, break down so often.

The same request, a different result each time

Picking the next piece, the model isn't holding just one candidate. It scores several plausible candidates by how well each one fits, then draws one from among them. A bit of randomness sits inside that draw, which is why the same request produces a different result every run.

How much randomness is allowed can be dialed. Narrow it and results stay safe and similar; widen it and unexpected combinations show up more often, at the cost of more awkward ones too. That randomness is exactly why running something several times until a result you like shows up actually works.

Only the material changes; the method stays the same

What gets invented can be text, a picture, a voice, or video. Change the material and the shape of a "piece" changes with it — a word fragment for text, one refining pass over the whole frame for a picture, a short slice of time for sound.

Pick one piece, attach it, repeat until done — that backbone stays the same. That's exactly why a technique built for text generation could carry over to pictures and sound, and why services covering several kinds of media at once have multiplied so quickly.

3More precisely

Generative AI refers to a model that learns the patterns in its training data as a shape of probability, then draws new results by following that shape. Where a purely judging model focuses on drawing a boundary line, a generative model learns the actual shape of how the data is spread out — which lets it pick out a point in that spread that nobody has ever made before.

The analogy breaks down in a few places. A florist reaches for real flowers sitting in a real bucket, while a generative model doesn't hold any finished sentence or picture inside it. The material it trained on isn't stored — every result is freshly computed. Still, something seen very many times in training can come out looking very close to it, which is exactly where copyright questions arise. And a florist knows for themselves whether the bouquet they tied looks good; a generative model has no such check — it only weighs plausibility, with nothing built in to confirm whether something is actually true. A florist also stops the moment the bouquet looks right, while a generative model keeps going until it has produced the full length it was asked for, whether or not an earlier piece was the best one to build on.

4Try it yourself

Craiyon (free image generation) ailearn.space Type in a single line and watch a new picture appear on the spot. Run the same line twice and compare the results

5Common misconceptions

It's easy to think it's fetching text or pictures stored somewhere, but actually it's freshly computed every time — the fact that the same request gives a different result each run is the proof.
It's easy to think something newly generated is also new and accurate as fact, but actually it's built to be plausible, so statements that don't match reality can slip in among perfectly natural-sounding sentences.
It's easy to think it creates the way a person does, but actually it's picking combinations from within patterns it learned, so a format with no precedent at all in its training tends to come out poorly.

7One-line summary

In shortGenerative AI invents a new result every time by following patterns it has learned, one piece at a time, instead of picking from a set of ready-made answers.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02