AI vs machine learning vs deep learning: the difference

Three nested circles, not three rivals

Picture three circles, one inside another. The outermost is artificial intelligence: the whole endeavour of making machines do things that would need intelligence from a human — understanding language, recognising images, making decisions, planning. Inside it sits machine learning: one particular approach to building AI, in which the system learns its own rules from data rather than having every rule written by a programmer. Inside that sits deep learning: one particular family of machine-learning techniques, built on neural networks with many layers.

So every deep-learning system is machine learning, and every machine-learning system is AI — but not the reverse. Early chess programs were AI without machine learning: humans hand-wrote their rules. A spam filter that learns from examples is machine learning without necessarily being deep learning. Generative AI, meanwhile, is best pictured as a recent, fast-growing branch sprouting from the deep-learning circle — the layered-network techniques applied to creating content rather than only classifying it.

Artificial intelligence: the umbrella term

AI is a goal, not a technique. Any system performing a task we would consider intelligent — parsing a spoken request, routing deliveries, spotting a tumour on a scan, playing a game well — counts as AI regardless of how it works inside. The oldest AI systems were entirely hand-crafted: experts encoded rules (“if the email contains these phrases, flag it”) and the machine followed them. That rule-based approach still runs quietly inside plenty of software today.

An everyday example of the umbrella at work is a satnav: it perceives where you are, plans a route, and re-plans when you miss the turn. Whether its internals are hand-written algorithms or learned models, the behaviour is what earns the AI label. Keeping “AI” this broad is precisely what makes the next two terms necessary — they say how the intelligence is achieved.

Machine learning: systems that learn from data

Machine learning flips the old approach. Instead of programmers writing the rules, you show the system many examples and it works the rules out for itself. A spam filter is the classic everyday case: rather than maintaining an ever-growing list of forbidden phrases, you feed it thousands of emails labelled spam or not-spam, and it learns which patterns distinguish them — adapting as spammers change tactics, without anyone rewriting code.

Machine learning comes in three broad flavours, each of which the AI Practitioner exam expects you to recognise at a conceptual level:

• Supervised learning — the system learns from labelled examples, mapping inputs to known correct answers, like the spam filter or a model predicting house prices from past sales.
• Unsupervised learning — the system finds structure in unlabelled data on its own, such as grouping customers into segments nobody defined in advance.
• Reinforcement learning — the system learns by trial and error, taking actions and improving from rewards and penalties, the way game-playing agents learn winning strategies.

Deep learning: many-layered neural networks

Deep learning is machine learning done with neural networks — models loosely inspired by the brain, built from layers of simple interconnected units. “Deep” refers to stacking many layers. Each layer transforms the output of the one before, so the network builds understanding progressively: in an image, early layers detect edges, middle layers combine edges into shapes and textures, and later layers recognise whole objects — a face, a stop sign, a cat. No human specifies those stages; they emerge from training on large amounts of data.

This mattered because some problems had resisted decades of rule-writing. Nobody can hand-write rules that reliably recognise faces at any angle in any light, or transcribe casual speech. Deep learning cracked exactly these perception-and-language problems, which is why the technology behind photo search, voice assistants and translation changed so dramatically within a few years. Today’s large language models and generative AI are deep learning at very large scale — the same layered-learning idea, pointed at producing text and images rather than only recognising them.

Why the terms get muddled

Marketing bears much of the blame. “AI-powered” sounds better on a product page than “contains a regression model”, so the grandest available term gets used for everything — including plain rule-based automation with no learning in it at all. Headlines compress further, using “AI” for whatever system is in the news that week. And because generative AI has dominated recent attention, many people now use “AI” to mean chatbots specifically, which inverts the actual hierarchy.

The muddle is harmless in casual conversation and costly everywhere else. In job adverts, “machine learning engineer” and “AI engineer” describe different work. In architecture discussions, choosing between a rules engine, a classical machine-learning model and a deep-learning model is a real decision with real trade-offs — classical models are often cheaper, faster and easier to explain, and remain the right choice for many business problems. Precision with these three words is a small habit that signals genuine understanding, to interviewers and examiners alike.

The hierarchy in certification study

If you sit the AWS Certified AI Practitioner (AIF-C01), this article is the syllabus’s opening move. The exam’s AI and machine-learning fundamentals expect you to define the three terms, place them in the correct nested relationship, distinguish supervised, unsupervised and reinforcement learning, and recognise which category a described scenario belongs to. Generative AI is then treated as its own topic area building on these foundations — the same shape as this article, which is no coincidence.

The AWS Certified Cloud Practitioner touches the hierarchy more lightly, expecting awareness that AWS offers AI and machine-learning services and roughly what problems they address. For either exam — and for reading the field’s news with a clear head — the nested-circles picture is the piece worth committing to memory: AI the goal, machine learning the data-driven approach, deep learning the layered technique, generative AI the newest branch.