Artificial intelligence

Your phone recognises your face, your spam filter sorts out annoying mails and your music app suggests songs you actually like. Behind all of this sits the same idea: machines that learn from examples. How that works is easier to understand than you might think.

Following rules or finding patterns

A classic program follows rules a human wrote down: if the price is above 10 euros, show a warning. For many tasks that works great. But try writing rules for how to recognise a cat in a photo. Nobody can. Learning systems take a different path: they get many examples, the so-called training data. That can be thousands of photos, each one labelled cat or dog. From these they work out the patterns that tell the two apart, all on their own.

Learning a boundary

Many learning tasks come down to the same thing: finding a boundary that separates two groups. The system starts with a random line, looks at example after example and nudges the line a bit with every mistake. After enough rounds, it sits so that even new, unseen examples usually land on the correct side. That is exactly what learning means here: not memorising, but generalising.

Why AI makes mistakes

A learned boundary is never perfect. Some points land on the wrong side, and what the system never saw in its training data, it cannot know. The data determines the quality: if it is one-sided or flawed, the machine learns exactly those flaws too. So check an AI's answers like the advice of a person who knows a lot but is sometimes wrong. You will still meet it daily: in face recognition, in spam filters, in translations, in navigation and in every recommendation feed.