System Atlas
the atlas

AI vs ML vs deep learning vs generative AI

Four nested circles that get thrown around interchangeably — untangled with one example each.

the gist

These four terms nest inside one another. Artificial intelligence is the broadest goal — any machine that mimics intelligence. Machine learning is the subset that learns from data instead of following hand-written rules. Deep learning is the subset of ML using many-layered neural networks. Generative AI is the subset of deep learning that creates new content. Every generative model is deep learning, is machine learning, is AI — but not the other way around.

ANALOGY // Russian nesting dolls: open 'artificial intelligence' and you find 'machine learning' inside, open that and you find 'deep learning', and inside the smallest sits 'generative AI'.
living diagram01 / 05
Each ring is a strict subset of the one above it
01
Artificial Intelligence
any machine that mimics intelligence — e.g. a rule-based chess engine
02
Machine Learning
learns patterns from data — e.g. a spam filter
03
Deep Learning
many-layered neural networks — e.g. image recognition
04
Generative AI
creates new content — e.g. ChatGPT, DALL-E
01

AI — the outermost goal

Artificial intelligence is the broadest term: any technique that lets a machine perform tasks we'd consider intelligent — reasoning, planning, understanding language, recognising images. Crucially it includes approaches with no learning at all, like a chess engine that follows hand-crafted rules and searches ahead. AI is the goal; everything below is a family of methods for reaching it.

02

Machine learning — learning from data

Machine learning is the subset of AI where, instead of a human writing explicit rules, the system learns patterns from data. You show a spam filter thousands of labelled emails and it figures out for itself which features signal spam, then generalises to new messages. The defining shift is from programming the answer to programming the ability to learn the answer.

03

Deep learning — many-layered neural nets

Deep learning is the subset of machine learning built on neural networks with many stacked layers. Its superpower is automatic feature learning: rather than an engineer hand-designing what to look for, each layer learns progressively more abstract patterns — edges, then shapes, then whole objects in an image. This is what unlocked breakthroughs in vision, speech and language once data and compute became abundant.

04

Generative AI — creating new content

Generative AI is the subset of deep learning whose job is to create new content rather than just label existing content. Where a classic deep model might answer 'is this a cat?', a generative model produces a brand-new cat image, a paragraph of text, or a snippet of audio. Large language models and image generators like DALL-E live here.

05

How they nest together

The four terms nest strictly inside one another: every generative model is deep learning, every deep learning model is machine learning, and every machine learning system is AI. The reverse never holds — a rule-based chess engine is AI but not ML, and a linear-regression spam score is ML but not deep learning. Placing any system in the right ring is the fastest way to cut through the buzzword soup.

the full picture
01

Why these four terms get tangled

"AI", "machine learning", "deep learning" and "generative AI" are used almost interchangeably in headlines, marketing and casual conversation — which is a shame, because they have a precise and genuinely useful relationship. They are not four competing things, and they are not synonyms. They are four concentric rings, each one a subset of the one outside it: generative AI sits inside deep learning, which sits inside machine learning, which sits inside artificial intelligence.

Getting this straight is more than pedantry. When someone says "we use AI", the word alone tells you almost nothing — it could mean a spreadsheet of hand-written rules or a frontier language model. Knowing which ring a system lives in tells you how it works, what it needs (data? rules? enormous compute?), and what it can and can't do. The clearest way to learn the relationship is to walk from the outermost ring inward, pinning down each term with a concrete example, and then see how the containment works.

02

Artificial intelligence: the outer ring

Artificial intelligence is the oldest and broadest term, dating to the 1950s. It refers to any technique that enables a machine to perform tasks that would normally require human intelligence — reasoning, problem-solving, perception, understanding language, making decisions. The key word is any: AI is defined by the goal, not the method.

That breadth is important because a huge amount of what has historically been called AI involves no learning whatsoever. A chess engine like the classic Deep Blue is AI: it plays a deeply intelligent game, but it does so through hand-crafted evaluation rules and brute-force search over future moves, all written by human programmers. Expert systems that encode a doctor's diagnostic rules as a giant if-then tree are AI. A game character that follows a fixed decision tree is AI. None of these learn from data — a human explicitly programmed the intelligent behaviour.

So the outer ring includes two broad families: symbolic or rule-based approaches, where humans write the logic, and learning-based approaches, where the machine infers the logic from data. That second family is the next ring in.

03

Machine learning: learning from data

Machine learning is the subset of AI in which the system is not explicitly programmed with rules but instead learns them from data. The paradigm shift is profound: rather than a programmer writing "an email is spam if it contains these 500 phrases", you show the system thousands of examples labelled "spam" or "not spam" and let it discover the predictive patterns itself, then apply them to emails it has never seen.

The classic example is exactly that spam filter, but the family is enormous: recommendation engines that learn your taste from your clicks, fraud detection that learns the shape of suspicious transactions, credit scoring, demand forecasting. These often use techniques like decision trees, linear and logistic regression, or support-vector machines — many of which are not "deep" at all.

The defining trade is that machine learning replaces hand-written rules with learned ones, which makes it far better at messy, high-dimensional problems where nobody could write the rules by hand — but it now depends utterly on data. Good, representative data yields a good model; biased or thin data yields a biased or brittle one. This is why "garbage in, garbage out" is the central anxiety of every ML project, and it's a concern the outer rule-based ring never had.

04

Deep learning: letting the network find the features

Deep learning is a subset of machine learning that uses artificial neural networks with many layers — the "deep" refers to that depth. Neural networks are loosely inspired by the brain: layers of simple interconnected units, each passing signals forward, with the connection strengths (weights) adjusted during training.

What makes deep learning special is automatic feature learning. In traditional machine learning, a human expert usually has to hand-engineer the input features — deciding, for an image, that the model should look at edges, colours and textures. Deep networks skip this: given raw data and enough of it, each successive layer learns to detect progressively more abstract features on its own. In image recognition the first layers learn edges, middle layers learn shapes and parts, and deeper layers learn whole objects like faces or cats — all discovered, not designed.

This capability is why deep learning triggered the modern AI boom. It excels precisely at the perceptual, unstructured problems — vision, speech, natural language — that stumped earlier methods, but it pays for that power with two demands: it needs large amounts of data and heavy compute (typically GPUs) to train. When data and hardware became abundant in the 2010s, deep learning went from a niche idea to the dominant approach, and it is the foundation everything in the innermost ring is built on.

05

Generative AI: from recognising to creating

Generative AI is the newest ring: the subset of deep learning that creates new content rather than merely analysing existing content. This is the crucial distinction. Most deep learning is discriminative — it takes an input and produces a label or number: is this a cat, is this transaction fraud, what's the sentiment of this review. Generative models instead produce brand-new outputs that resemble their training data: a novel image, an original paragraph, a piece of music, a working function of code.

The examples are the ones dominating the headlines. Large language models like the one you're reading about generate text token by token. Diffusion models and GANs like DALL-E, Midjourney and Stable Diffusion generate images. There are models that generate audio, video and 3D shapes. All of them are deep neural networks — usually very large ones — trained on massive datasets, which is exactly why generative AI sits firmly inside the deep learning ring.

What changed with generative AI is less the underlying mechanism than the interface and capability: for the first time, non-experts can interact with a powerful deep-learning model in plain language and get creative, open-ended output. That accessibility is why "AI" in everyday 2020s conversation usually means generative AI specifically — even though it's only the innermost slice of a much larger field.

06

Why the nesting matters in practice

The single most important property of this hierarchy is that the containment runs one way only. Every generative AI system is also deep learning, is also machine learning, is also AI. But the reverse fails at every step: a rule-based chess engine is AI but not machine learning; a decision-tree fraud detector is machine learning but not deep learning; an image classifier is deep learning but not generative. The inner rings are strict specialisations of the outer ones.

Why does this matter beyond tidiness? Because placing a system in the right ring immediately tells you a lot about it:

  • What it needs — rule-based AI needs human-written logic; ML and below need data; deep learning and below need lots of data and compute.
  • How it fails — rule-based systems fail on cases nobody wrote a rule for; learning systems fail on data they weren't trained on, and can absorb the biases in that data.
  • Whether it's the right tool — a simple, well-understood problem may need only a handful of rules or a small ML model, and reaching for a giant generative model would be wasteful and harder to trust.

So when you next hear a product "uses AI", the useful question is: which ring? A hand-coded rule, a learned classifier, a deep network, or a generative model? The answer changes what you should expect, what could go wrong, and how much data and compute stand behind it. The buzzwords collapse into clarity the moment you see them as nested circles rather than rival labels.

what to remember
01

The relationship is strictly nested: AI contains ML contains deep learning contains generative AI.

02

AI is any technique that makes machines mimic intelligent behaviour — including old-school hand-coded rules.

03

ML is the subset where the system learns patterns from data rather than being explicitly programmed.

04

Deep learning is ML using neural networks with many layers, which learn features automatically.

05

Generative AI is deep learning that creates new content (text, images, audio) rather than just classifying it.

06

The nesting is one-directional: ChatGPT is all four; a rule-based chess engine is only AI.

check yourself

Quick quiz

Q1

Which statement about the nesting is correct?

Q2

A chess engine that follows only hand-written rules is…

follow the thread