How AI actually works, without the maths
AI isn’t programmed rule by rule. It learns from examples, spots patterns and makes predictions. Here’s how that works, why it improved so fast and where it slips up.
Video transcript
Your phone can find every photo of your dog, yet nobody ever told it what your dog looks like. So how does it know? Here’s how AI works, with no maths.
Take your spam filter. Old-style software follows rules a person writes by hand. But scammers keep changing their wording, so no list of rules could ever catch them all.
So instead, you show the computer thousands of emails, each marked spam or not spam, and let it work out the warning signs for itself. That’s machine learning.
The learning stage is called training. It guesses, checks the answer and adjusts, millions of times. What comes out is a model, which makes predictions about emails it has never seen.
Why the big leap recently? Three things came together: far more data to learn from, far more computing power, and better methods, like neural networks, loosely inspired by the brain.
But here’s the catch. AI spots patterns. It doesn’t understand the world the way you do. And it learns whatever is in its examples, so if they’re unfair, its answers can be unfair too.
That’s AI in one line: examples in, predictions out. Read the full guide below, then try asking an assistant to explain how it was trained.
In 30 seconds
- AI learns from examples, not hand-written rules: training turns them into a model that makes predictions.
- It improved fast thanks to far more data, far more computing power and better methods.
- It spots patterns rather than understanding them, and it can pick up the biases in its examples.
Search your phone’s photos for “dog” and up pop pictures of your dog, even though you never labelled a single one. Nobody wrote instructions describing your dog, either. The app learnt what dogs look like from examples, and that idea sits behind almost all of today’s AI.
Rules versus learning
Ordinary software follows an Jargon busterAlgorithm: A set of step-by-step instructions a computer follows. that a programmer writes out in full: if an email mentions a lottery win, send it to spam. That works until the scammers change a word. You can’t write a rule for every scam email, because there are endless ways to word one.
So instead, you show the software thousands of emails, each marked “spam” or “not spam”, and let it work out the warning signs for itself. This is called Jargon busterMachine learning: The branch of AI where software learns from examples instead of being programmed rule by rule., and it’s how almost all modern AI is built.
| Old way: rules | AI way: examples |
|---|---|
| A programmer writes every instruction | The software learns from examples |
| “If it says ‘lottery’, it’s spam” | It spots what scam emails have in common |
| Breaks when scammers change a word | Can catch new scams that look like old ones |
| Does exactly what it’s told | Makes its best guess, so it can be wrong |
Training, models and predictions
The learning stage is called Jargon busterTraining: The stage where AI learns, by finding patterns in a huge number of examples.. The software looks at an example, makes a guess, checks the right answer and adjusts itself very slightly. Then it does it again, millions of times. Nobody tells it what to look for.
What comes out is a Jargon busterModel: What training produces: a very large set of numbers that captures patterns, used to make predictions.: everything it learnt, stored as a huge set of numbers. Show the model something new, such as an email that has just arrived, and it makes a prediction: probably spam.
- ExamplesThousands of emails, marked spam or not spam
- TrainingGuess, check, adjust, millions of times
- ModelWhat it learnt, stored as numbers
- PredictionA new email arrives: “probably spam”
Why it suddenly got so good
Machine learning isn’t new: spam filters have used it for years. What changed is that three things came together.
- Far more data. The internet put billions of photos, posts and pages within reach, so there were examples of almost everything.
- Far more computing power. Chips first designed for video games turned out to be ideal for training, so models could learn from far more examples.
- Better methods. A design called a Jargon busterNeural network: A kind of model loosely inspired by the brain: layers of simple maths units that pass numbers to each other. proved excellent at finding patterns, especially once it was made much bigger.
Chat assistants were built the same way, from text rather than photos: here’s what a large language model is.
Patterns, not understanding
Here’s the catch. AI is very good at spotting patterns, but it doesn’t understand the world the way you do. Your photos app can find your dog in seconds without any idea what a dog actually is.
It also learns whatever is in its examples, flaws and all. If most of the dogs it saw were Labradors, it may miss your whippet. If past hiring decisions favoured men, a system trained on them can learn to favour men too. That’s what people mean by bias in AI.
Try it yourself 2 minutes
- Open the photos app on your phone and search for “dog”, “beach” or “cake”.
- Look at what it found, and at what it missed or got wrong.
- The misses show where its patterns run out, such as odd angles or bad light.
Explain how you were trained, as if I’m 12 years old. Use one everyday comparison and keep it under 150 words. Then name one thing you’re bad at, and why.
So AI is a pattern-spotter, not a know-it-all. That’s why it can sound sure and still be wrong: here’s why AI makes things up, and how to catch it.
Check yourself
3 quick questions nothing is savedTools in this guide
Spotted a mistake? Tell us and an editor will check it.
