What is machine learning?

You tap your card at a coffee shop in a town you've never visited. A minute later your phone buzzes. Did you make this purchase? Nobody at the bank was watching your account. Software flagged it in seconds, and very often that software is built a particular way, called machine learning. It's one of the main ways people build artificial intelligence (AI), and once you see how it works, a lot of AI starts to make sense.³

For a long time, a system like that ran on rules a person wrote. Flag any purchase over a certain amount. Flag any purchase in another country. Rules like those are easy to read, but they're also easy to learn. A fraudster can keep each purchase small and local and slip under every line. And every new trick means someone has to spot it first, then write another rule.³

Machine learning turns that around. Instead of writing the rules, people give the computer examples: millions of past purchases, each one marked as fraud or fine. The computer works through them and finds the patterns that tend to separate the two. Then, when a new purchase comes in, it uses those patterns to make a guess. Put plainly, machine learning is software that learns its rules from examples, instead of having every rule written by hand.¹²

The rule it finds may be one no person could list. It might weigh the time of day, the kind of shop, how far you are from your last purchase, and how much you usually spend, all at once, in ways that would take pages to write out. Nobody typed that rule in. It came out of the examples. That's the part people find surprising, and it's the heart of the idea: the people who build these systems choose the examples and check the results, and the computer finds the rule.¹²

Fraud keeps changing, so the examples change too. When new tricks show up in the records, the system can be trained again on fresher examples, and it can pick up patterns that weren't there before. Nobody has to guess the next trick in advance. People still decide when to retrain it, and they check that the new version does better than the old one before it goes to work.³

There's a catch, and it explains a lot about AI. A system that learns from examples is only as good as those examples. If most of what it learned about you happened close to home, a coffee in a new town can look odd, even when it's you. That's why a trip can set off a false alarm. Nothing broke inside it. It did what it learned, from the examples it was shown. And it's one reason a text like that asks you to confirm, so the person who knows where they are gets the last word.¹

You'll find the term on product pages, in news stories, and in job ads. Now you can read it plainly: somewhere, a computer learned a rule from a pile of examples. And when a tool tells you it uses machine learning, you can ask the most useful question there is about it: what examples did it learn from?