What Is AI, Really?
By the end of this lesson
Distinguish AI, machine learning, deep learning and generative AI, and set accurate expectations.
These four terms get used as though they were synonyms. They are nested, and knowing which one you mean prevents a surprising amount of muddled thinking about what a system can be expected to do.
- Artificial intelligence
- The broad field of getting software to do things we associate with human intelligence. A chess program from 1990 counts. It is a field, not a technique.
- Machine learning
- An approach within AI where behaviour is derived from data rather than written as rules. You supply examples; the system finds patterns. Nobody writes "if the email says 'free money', mark it as spam" — the pattern is learned.
- Deep learning
- Machine learning using neural networks with many layers. This is what made speech recognition, image recognition and language models practical.
- Generative AI
- Deep learning models that produce new content — text, code, images — rather than classifying existing content. A spam filter decides; a language model writes.
So: generative AI is a kind of deep learning, which is a kind of machine learning, which is a part of AI. When someone says "we should use AI for this", the useful follow-up is which of these they mean, because the answer determines whether the problem is even a fit.
Rules versus learned patterns
The genuine dividing line for deciding what to build:
| Write ordinary code | Consider machine learning | |
|---|---|---|
| The rules are | Known and expressible | Hard to state, but examples are plentiful |
| Example | Calculate tax owed on an invoice | Decide whether a review is positive or negative |
| Correctness | Exactly right or wrong | Statistically good, occasionally wrong |
| Explaining a result | Straightforward | Often difficult |
Setting accurate expectations
What generative models are reliably good at:
- Rewriting, summarising and changing the tone of text
- Pulling structured information out of unstructured text
- Classifying content into categories you describe
- Drafting code and explaining unfamiliar code
- Answering questions when you supply the relevant material
What they are not reliable at:
- Arithmetic and precise calculation
- Recalling specific facts accurately, particularly recent ones
- Producing identical output for identical input
- Knowing what they do not know — they rarely decline, they guess
- Anything requiring an audit trail of how a conclusion was reached
Summary
- AI contains machine learning, which contains deep learning, which contains generative AI
- Use ordinary code when the rules are known; consider a model when they resist being written down
- Models are strong on language-shaped tasks and weak on arithmetic, recall and repeatability
- Confident wrong answers follow from how the technique works, so supply facts and validate output
Practice
Attempt each one before opening the solution. Getting it wrong first is how the idea sticks.
Think about it
Think about it
For each of these, decide whether you would write ordinary code or use a language model: calculating a shipping cost from weight and destination; sorting incoming support emails by topic; checking whether a postcode is valid; drafting a reply to a customer complaint.
Show solution
Shipping cost: ordinary code. The rules are known and must be exact.
Sorting emails by topic: a model is a reasonable fit. The input is free text and the categories are fuzzy.
Postcode validation: ordinary code. It is a format rule, and a model would be slower, costlier and less reliable.
Drafting a reply: a model suits this well, with a human reviewing before it is sent.
The pattern: exact rules and exact answers mean code. Messy input and tolerance for review mean a model.
Knowledge check
Nothing is recorded and there is no score. The explanation appears either way.
Saved in this browser only.