AI, machine learning, deep learning and generative AI

Artificial intelligence, machine learning, deep learning and generative AI are often used as if they meant the same thing. Each one sits inside the one before it.

You type a question for a class into an AI chatbot. A clear paragraph appears within seconds, in full sentences. Other software you use every day, such as a spam filter, a map app or a photo app, is built in different ways, and the four terms describe those ways.

Your questionTyped text
The chatbotLanguage model
The answerGenerated text
Think first. Did people write down the rules this chatbot follows, or did the program work out its own rules from examples? Decide, then check in the next step.
Ask for a show of hands before students open the next step. Most will say the chatbot learned from examples. Follow up with the chess engine and the route finder, which learned nothing. The rule-based case matters most for the lessons that follow, because they assume students can tell a learned model from a hand-written program.

Select a system in the list to see which circles it falls inside. Two of the eight are AI that does no machine learning at all.

Choose a system

Read the circles from the outside in.

Artificial intelligence

Any program that does something we would call intelligent. Most of it is hand-written rules.

Machine learning

Programs that work out their own rule from labeled data.

Deep learning

Machine learning with stacked layers that choose their own features from raw input.

Generative AI

Systems that produce new content, such as text, images or code.

More detail

Examples of each kind

Artificial intelligence: chess engines, route finding, scheduling. Machine learning: spam filters, credit scoring, recommendations. Deep learning: face recognition, speech to text, translation. Generative AI: language models, image models, coding assistants.

Labeled data and features

For a spam filter, labeled data is a million emails already marked spam or not spam. A feature is one measurement a person decides to feed the program, such as the number of links. Deep learning takes over the job of choosing features, so you give it the raw input, such as pixels or sound.

Generative AI and the other three

A spam filter gives one of two answers. A generative system writes out something that did not exist before.

The term AI dates from 1956, and much of what runs banks is hand-written rules. Today a product that says "AI" usually means generative AI. Ask which kind it is. A hand-written rule gives the same wrong answer every time for the same input. A trained model's mistakes depend on the examples it was trained on.

You can now tell which kind of AI a product uses.

More detail

Most AI does not learn

A great deal of what runs banks, airlines and traffic lights is rules that people wrote. It works, and nobody markets it as AI any more. What changed recently is the small innermost circle in the diagram.

Where the lessons go from here

The lessons that follow are about the innermost circle. It covers what the model predicts, how it is trained, what it reads, and what happens when you give it tools.