What AI actually is, and what it is not

Most explanations of AI are either frightening or meaningless. Here is the plain version.

Ordinary software follows rules a person wrote. A calculator adds because someone wrote the addition. A fee-collection screen marks a student overdue because someone wrote 'if the date has passed and the money has not arrived'. Every behaviour was decided in advance by a human.

AI is software that works out the rules from examples instead. Nobody wrote down what a cat looks like. They showed a system a very large number of pictures labelled cat and not-cat, and it worked out the pattern by itself.

That single difference explains almost everything else about AI — including its weaknesses.

Because it learned from examples:

- It is very good at things where good examples exist in huge numbers.

- It is unreliable where they do not.

- It cannot tell you *why* it decided something, in the way a rule could.

- It absorbs whatever was in the examples, including the mistakes and the bias.

Three words you will hear used loosely:

AI — the whole field. Often used to mean nothing in particular, in advertising.

Machine learning — the actual method: learning patterns from data. Nearly all working AI is this.

Generative AI — the recent kind that produces new text, images or code. ChatGPT is this.

So when someone says 'we added AI', the useful question is: *learned from what examples, to do what exactly?* If they cannot answer, there is often nothing there.