The relationship in one sentence
Machine learning is one approach to AI. Deep learning is one technique within machine learning. Generative AI is what we use that technique for most spectacularly today.
Going outwards, each is broader than the one before it. Going inwards, each is narrower and more specific. That is the whole relationship — everything else is detail.
1. Machine learning: the machine works out the rule
In traditional programming you write the rule yourself: "if the email contains the word prize, put it in the spam folder." That works until the spammer rewrites the word as pr1ze.
Machine learning reverses the direction. You do not supply a rule, you supply examples: a hundred thousand emails, each labelled spam or not spam. The system finds out for itself what separates the two. It does not only look at words — it looks at the sender's history, the ratio of links, the time of day, correlations you would never have thought to write down.
None of this is new or mystical. Your bank's fraud screening, a shop's recommendation engine and demand forecasting all work this way, and have done for the best part of twenty years. These systems create nothing new: they classify, forecast and recommend. Today we call them predictive systems, and at many companies they are still what delivers the most measurable return.
2. Deep learning: many layers, one after another
Early machine learning ran into a wall at one particular point: someone had to tell the machine what to look at. If you wanted it to recognise a cat, a specialist had to write down that cats have pointed ears and whiskers, and those characteristics had to be extracted from the image separately. This was called feature engineering, and it was the harder half of the job.
Deep learning hands that step to the machine as well. The model is built from many layers — sometimes hundreds — and the layers learn stacked abstractions on their own: the first layer finds edges, the next shapes, the next ears and eyes, the top one decides "this is a cat." Nobody told it that ears matter. It emerged from the examples.
So "deep" does not mean it has deep thoughts. It means many layers.
The technique became usable around 2012, and not because of a theoretical breakthrough: the internet had produced enough data, and graphics cards had become cheap enough. The why-now question gets a fuller answer in What AI Really Is.
3. Generative AI: producing rather than sorting
Up to this point every system turned input into a decision: spam or not spam, cat or dog, 300 or 500 units sold.
The output of generative AI is not a decision but content: a sentence, an image, an audio track, a snippet of code. Today's models also handle several kinds at once — that is the subject of Multimodal AI. The same deep learning technique sits underneath, just pointed at a different job: instead of predicting a label, it predicts what comes next. In a language model that is literally what happens, which is what How a Machine Guesses the Next Word is about.
This shift became visible in 2022, and it is what everyday speech now means by AI.
What the diagram does not show
Two misunderstandings are worth stating out loud.
Generative is not "more advanced" than predictive. It is for a different job. If you want to know how many units to order for next month, a predictive model gives you a more accurate, cheaper and more explainable answer than any chatbot. Plenty of companies see no return from AI because they picked the impressive tool rather than the right one.
Deep learning is not the only machine learning. On tabular business data — revenue, customer age, order counts — twenty-year-old methods built on decision trees still routinely beat neural networks. They learn from less data, run faster, and let you inspect afterwards what drove the decision. In a credit assessment that last point is not a nicety, it is a legal requirement.
One example that shows all three
Take an online shop.
- Predictive, not deep: which customer is likely to churn? A few dozen columns, a decision tree. Runs on a laptop, and tells you which three factors mattered.
- Deep, not generative: which category does the product in this uploaded photo belong to? Image recognition, a many-layered network.
- Generative: write a product description from the photo and the specifications. A language model.
Same shop, three different technologies. Calling all three "AI" is not wrong — but using the third to solve the first burns money.
What to take away
Next time someone talks about AI, ask yourself one question: is it sorting or producing? That question cuts the field in half, and the answer tells you what data it needs, what it costs, and how far you can trust it.