How AI, ML, Deep Learning and GenAI Fit Together

AI, ML, deep learning, GenAI. You hear these words all the time. They're not four separate things: each one sits inside the one before it. Scroll to zoom in, one ring at a time.

Every ring is one kind of the ring around it.

Ring 1 · The outer circle

AI is the goal

Artificial intelligence means a machine doing a task we'd call smart: sorting, recognising, deciding, creating. The word names the goal, not the method.

Two robots sort a crate of fruit into apple, banana and green apple bins. One follows a rule card; the other learned from example fruit. Both succeed. Takeaway: AI is the goal. Rules and learning are two ways to build it.

Rules: people write every step. "Red and round → apple."
Learning: the machine works out the pattern from examples.
Both count as AI. Most AI you use today takes the learning route.

ThinkA rule-based sorter meets a fruit nobody wrote a rule for. What happens?›

It gets stuck or sorts it wrong: it can only do what its rules say. A sorter that learned from lots of varied examples can often cope. That flexibility is why most modern AI learns.

Like two cooks making the same dinner: one follows a recipe card, the other learned by tasting. Both are cooking.

The learning route has a name. Step inside: Machine learning ↓

Ring 2 · Inside AI

Machine learning learns from data

Instead of writing rules, you show the machine examples that already have the right answer. It finds the pattern, and that pattern is the model.

Train: house cards with prices become dots on a size versus price chart with a fitted line. Predict: a new house becomes a purple dot on the line, which points to its price tag. Takeaway: different data, different model.

Train: show it past houses, each with its size and sale price.
Learn: it finds the line that fits those dots best. That line is the model.
Predict: give it a new house size, and it reads the price off the line.

It only knows its dataTrain it on Perth houses and it's wrong about Sydney. Never shown a mansion? It guesses badly. The data decides the answer.

ThinkYou train it only on small apartments, then ask about a big farmhouse. Should you trust the price?›

No. The farmhouse is outside anything it has seen, so the line is just stretching into a guess. Train it on different houses and you'd get a different line. Different data, different model.

Like an agent who learned house prices by watching hundreds of sales, not by reading a rulebook.

A line works when the input is a couple of numbers. What about a photo? Step inside: Deep learning ↓

Ring 3 · Inside ML

Deep learning finds its own clues

A house is two numbers. A photo is millions of pixel numbers. Nobody can hand-pick the clues in all that, so deep learning stacks many layers that find the clues themselves.

A bicycle photo passes through three glass layers: the first shows edges, the second shapes like a wheel and frame, the third a whole bicycle, ending in a BICYCLE label. Takeaway: deep learning is many layers that find their own clues.

Early layers spot simple things: edges and lines.
Middle layers combine those into shapes: a wheel, a frame.
The last layer decides: "bicycle."
Shown thousands of labelled photos, all the layers learn together. Nobody programs what each one looks for.

Same rule as MLIt only knows what it has seen. Trained on bikes, cars and buses, it will call a horse one of those.

Look insideWhat does a neural network actually look like?›

Inside a neural network: a pixelated bicycle image feeds a column of input dots, three hidden layers of connected dots, and an output that picks BICYCLE over CAR and BUS.

Each dot is a simple unit that passes a signal on. Each line is a connection with a strength learned during training. "Deep" just means many layers of these.

ThinkWhy not just write rules for what a bicycle looks like?›

Bikes come in every colour, size, angle and light. The rules would never end. Letting the layers learn from examples works far better. This is exactly where deep learning shines.

Like an assembly line: each station adds a little more until a finished product rolls out.

So far every ring labels things. The last ring flips that. Step inside: GenAI ↓

Ring 4 · The centre

GenAI makes something new

Deep learning looks at a bicycle photo and says "bicycle." Generative AI goes the other way: ask for "a bicycle on the moon," and it makes a picture that never existed.

Left, faded: an ordinary bicycle photo with a LABEL tag, marked versus. Right: the prompt A bicycle on the moon goes to a robot built from stacked layers, which produces a new image of a bicycle on the moon. Takeaway: GenAI creates new content instead of labelling.

Built on deep learning: the same kind of layered network, trained on huge numbers of examples.
It learns patterns (what bikes look like, what the moon looks like), then combines them into something new.
It can make text, images, audio or code. Chat assistants are GenAI for text.

ThinkIs GenAI just copying a picture it found somewhere?›

No. It builds a new picture from patterns learned across millions of examples. That's why it can make things nobody has drawn, and also why it sometimes gets details wrong.

Like a musician improvising a new tune from years of songs they've heard.

The whole picture

AI: the goal. A machine doing smart tasks, by rules or by learning.
Machine learning: learns the pattern from data.
Deep learning: many layers learn their own clues.
GenAI: uses those layers to create.

Remember: each ring is one kind of the ring around it. Today's GenAI is built with deep learning, which is a kind of machine learning, which is one way to do AI.