How does AI actually work?

Artificial intelligence is not a single invention. It is a collection of techniques for getting computers to perform tasks that normally require human judgement: recognising a face, translating a sentence, recommending a film or predicting the next word in a paragraph.

Most of the AI you encounter today is machine learning. Instead of being given explicit rules, a program is shown many examples and adjusts itself to spot patterns in them.

Learning from examples

Imagine teaching someone to recognise spam email. You could write rules — "messages containing the word prize are suspicious" — but spammers change their language and the rules break. A machine-learning system takes a different approach: it is given thousands of labelled emails, spam and not-spam, and gradually tunes its internal settings until its predictions match the labels.

That tuning process is called training. The system makes a prediction, measures the error, and nudges its settings to reduce that error. Repeat this millions of times and the model becomes useful at the task.

Neural networks

Many modern systems use neural networks, which are loosely inspired by the brain. A network is organised into layers of simple units connected by adjustable weights. Data passes through the layers, and each layer transforms it a little.

In an image recogniser, early layers might detect edges, middle layers detect shapes, and later layers detect objects. Nobody programs those features directly; they emerge from training.

Deep learning simply means a neural network with many layers. It became practical in the 2010s because of three things: much more data, much faster graphics processors, and better training methods.

Large language models

A large language model (LLM) such as the ones behind modern chatbots is trained on enormous amounts of text. Its core task is almost embarrassingly simple: predict the next token (roughly, a word or word fragment) given the tokens before it.

Doing that well requires learning grammar, facts, styles of writing, common reasoning patterns and much more. When you ask a question, the model generates a response token by token, choosing likely continuations based on its training and the conversation so far.

This is why LLMs can sound fluent and confident — and also why they can invent plausible-sounding falsehoods. They are optimised for likely text, not for truth.

What AI is good and bad at

AI is strong at pattern matching, summarising, drafting, classifying and generating variations. It is weak at verifying facts it has not been given, understanding consequences, and knowing when it is out of its depth.

It is also shaped by its training data. Biases, gaps and outdated information in that data can show up in outputs.

Where you meet it

• Search and recommendations: ranking results, suggesting videos, filtering spam.
• Language tools: translation, transcription, summarising, drafting.
• Images and media: photo editing, generation, upscaling, moderation.
• Everyday software: autocomplete, fraud detection, route planning, voice assistants.

The useful question is not "is this AI?" but "what is it optimising for, what data did it learn from, and how would I know if it was wrong?" That habit will serve you better than any amount of jargon.