AI, machine learning, deep learning and generative AI: how they fit together

Many questions on the AWS Certified AI Practitioner (AIF-C01) exam rest on four terms: artificial intelligence, machine learning, deep learning and generative AI. They're often used as if they mean the same thing. They don't, and the exam expects you to know the difference.

This post covers material from Domain 1 (Fundamentals of AI and ML, 20% of scored questions) and the start of Domain 2 (Fundamentals of Generative AI, 24%). The domain weightings come from AWS's exam guide.

One picture: circles inside circles

The easiest way to remember how the four terms relate is as nested circles:

• Artificial intelligence (AI) is the outer circle: any technique that lets a computer do something we'd normally think needs human intelligence, such as understanding language, recognising images or making decisions. A hand-written set of rules can count as AI.
• Machine learning (ML) sits inside AI. Instead of being given rules, an ML model learns patterns from data. Show it enough examples and it can make predictions about new ones.
• Deep learning sits inside ML. It uses neural networks with many layers, which are especially good at messy, unstructured data like images, audio and free text.
• Generative AI sits inside deep learning. Its models create new content (text, images, code, audio) rather than only classifying or predicting. Most are built on foundation models: very large models trained on broad data that can be adapted to many tasks.

So every generative AI system is a deep learning system, every deep learning system is machine learning, and all of it is AI. The reverse isn't true.

The three main types of machine learning

Exam scenarios often describe a problem and ask which kind of learning fits. The clue is almost always in the data.

Type	What the data looks like	Typical tasks
Supervised learning	Labelled: each example comes with the right answer	Classification (is this email spam?), regression (what will this house sell for?)
Unsupervised learning	Unlabelled: no answers attached	Clustering (group customers by behaviour), anomaly detection
Reinforcement learning	No dataset; an agent learns from rewards and penalties as it acts	Robotics, game playing, optimising a sequence of decisions

A quick way to tell them apart in a question:

• "We have thousands of past transactions, each marked fraudulent or not" → supervised.
• "We want to find natural groupings in our customers, but we don't know what they are yet" → unsupervised.
• "The system should learn by trial and error to maximise a score" → reinforcement.

Classification or regression?

Within supervised learning, the exam also checks you can tell these two apart:

• Classification predicts a category: spam or not spam, which product type, which language.
• Regression predicts a number: a price, a temperature, a sales figure.

If the answer is a label, it's classification. If it's a quantity on a scale, it's regression.

When AI isn't the right answer

Domain 1 also asks you to recognise when AI isn't a good fit. Watch for scenarios where:

• the answer must be exactly right every time and a simple rule or calculation would do the job;
• there isn't enough good data to learn from;
• the decision has to be fully explainable and a simple, transparent method would meet that need better;
• the cost of building and running a model outweighs the benefit.

A model makes predictions, and predictions can be wrong. If a fixed rule solves the problem, the rule is usually the better choice.

How generative AI is different

Traditional ML models are usually built for one task: this model detects fraud, that one forecasts demand. Foundation models flip that around. One large model, trained once on a huge range of data, can summarise documents, answer questions, write code and draft emails, steered by the prompt you give it.

That flexibility comes with trade-offs the exam returns to again and again:

• Hallucinations: the model can produce confident, fluent answers that are wrong.
• Non-determinism: the same prompt can give different answers.
• Cost and latency: bigger models are slower and more expensive to run.

Knowing these trade-offs is what Domain 3 builds on, when you choose between prompting, retrieval-augmented generation and fine-tuning.

Test yourself

Can you answer these without scrolling up?

1. A model that writes product descriptions from a few bullet points: which of the four circles is the innermost one it belongs to?
2. A bank has years of loan applications, each marked "repaid" or "defaulted". Which type of learning fits?
3. Predicting tomorrow's electricity demand in megawatts: classification or regression?

(Answers: generative AI; supervised learning; regression.)

Want to know which domains you're already strong in? Take the free diagnostic, or go straight to the courses.

CertBrain is an independent study resource. It isn't affiliated with or endorsed by Amazon Web Services. AWS and AWS Certified AI Practitioner are trademarks of Amazon.com, Inc. or its affiliates.