What artificial intelligence is (and what it is not)

Over the past three years the letters AI have been stuck onto so many products that they stopped meaning anything. Your spam filter is AI. A film recommendation is AI. A chatbot that writes poetry is AI too. If that feels confusing, you are not alone — and it is not your fault.

One sentence that covers it

Artificial intelligence is software that learned its behaviour from data instead of having it programmed in, rule by rule.

That is the whole difference. And it is bigger than it sounds.

A classical program is a recipe. A developer writes: if the total is over a thousand euros and the customer is new, ask for payment up front. When the rule changes, someone rewrites it. The program never does anything its author did not tell it to — which is simultaneously its greatest strength and its hardest limit.

A machine learning model gets examples instead of rules. Tens of thousands of orders where we know which ones went unpaid. The model finds the patterns that separate them on its own. Nobody told it that the hour of the order matters, or whether the billing and delivery addresses match. It worked that out itself.

Why anyone bothers

Because some rules cannot be written down.

Try defining exactly what makes a photo of a cat a photo of a cat. Not "four legs and a tail" — a dog has those. Not "whiskers" — often you cannot see them. A person recognises it in a tenth of a second and cannot explain how. That is precisely where classical programming hits a wall and learning from examples works.

The same goes for speech, handwriting, the tone of an email, or guessing whether a customer is about to leave for a competitor. These are tasks where we have intuition but no rules.

What AI is not

It is not thinking. A model has no intentions, no opinions, no understanding of the world in the sense you have one. When a language model writes that it understands you, that is a statistically likely response to your sentence — not empathy.

It is not knowledge either. A model does not remember facts the way an encyclopedia does. It remembers patterns. That is why it can fluently write something that sounds right and is nonsense. This is called hallucination, and it is not a malfunction — it follows directly from how the model works.

And it is not magic. Behind every impressive output are the data it learned from. Where the data had nothing, the model fills the gap itself.

Where the line runs

This is probably the most useful thing to take away. AI is good at tasks where:

• plenty of examples exist to learn from,
• an occasional mistake is not a catastrophe,
• and a human can check the result.

It is bad — or at least risky — anywhere you need to be right every time, where a mistake actually costs something, and where nobody reviews the output. Leave VAT calculation to ordinary code. It does not get things wrong, and when it does, you know exactly why.

In practice it usually looks like this: AI does eighty percent of the work and a person does the rest. Less impressive than promises of full automation, but it works.

Why now

Neural networks are not a new idea; the foundations date to the 1960s. Three things changed at once: enough data (the internet), enough compute (graphics cards), and — since 2017 — an architecture called the transformer that learns far more efficiently than anything before it. That is what the language models everyone now knows are built on.

So no, no miracle happened. Conditions that had been missing for decades simply arrived together and finally got an old idea moving.

The takeaway

Next time somebody offers you an "AI solution", ask two things: what data did it learn from, and what happens when it gets it wrong. If the first question gets no concrete answer and the second gets "it won't", you are looking at marketing, not technology.