Most explanations of artificial intelligence start with a definition that sounds impressive and explains nothing. Here is one you can actually use.
Artificial intelligence is software that performs tasks we would normally say require human intelligence — recognising a face, understanding a sentence, planning a route, writing a paragraph.
That definition is deliberately loose, because the field is. It is a label for a goal, not for one technique.
The distinction that matters
For most of computing history, software did exactly what someone told it to do:
If the temperature is above 30 degrees, turn on the fan.
A person decided the rule. The computer applied it. If the situation changed, a person changed the rule.
Machine learning inverts this. Instead of writing the rule, you show the system thousands of examples and it derives the rule itself:
Here are 50,000 photos labelled "cat" and 50,000 labelled "not cat". Work out the difference.
Nobody can write down the rule for "this image contains a cat" — not in a way that survives contact with a blurry photo of a cat behind a chair. But a system given enough examples can find a pattern that works remarkably well.
This is the real dividing line. Traditional software encodes a human's understanding. Machine learning derives its own, from data.
Why the word covers so much
"AI" gets applied to systems that have almost nothing in common:
- A spam filter that learned which emails you delete
- A model that generates an image from a description
- A chess engine that searches millions of positions
- A chatbot that writes an email for you
Some of these learn from data. Some are pure search. Some are a thin wrapper over a rule someone wrote in 2004.
This matters practically. When someone tells you a product is "AI-powered", the useful follow-up question is not how advanced is it but what did it learn from, and what happens when it is wrong?
What this means for you
Through the rest of this course, when you meet a system described as AI, hold it to three questions:
- What does it actually do? Predict a number, classify something, generate text?
- What was it trained on? Whose data, from when?
- How does it fail? Not if — every one of these systems is wrong sometimes.
You will find that the systems worth trusting are the ones whose builders can answer all three.