How does AI work? Language models explained in plain words

The basic idea: predicting text

You may have noticed that your phone keyboard suggests the next word. Type "Happy" and it offers "birthday" or "Christmas". A chat AI such as ChatGPT, Gemini, Copilot or Claude works on the same principle – just vastly more skilfully.

When you type a question, it does not fetch a ready-made answer from an archive. It works out which word (or piece of a word) is most likely to come next, adds it, and then works out the next one. That is how a sentence, a paragraph and finally a whole answer appear – one word at a time, very quickly.

That sounds simple, but once the predictor has seen an unimaginable amount of text, it also picks up grammar, styles, facts, reasoning and even humour along the way. That is why its answers feel sensible.

What is a "model"?

When people talk about AI you often hear the word model or language model. The model is the part of the AI that has learned – a kind of compressed memory of how language works. Technically it is a huge set of numbers that were tuned during training.

You can picture the model as an experienced writer who does not remember individual books word for word, but has read so much that they can write convincingly about almost anything. ChatGPT, Gemini and Claude are apps with a company's own language model working inside them. The companies release new versions of their models regularly, and newer ones are usually more capable.

How is a model trained?

Roughly speaking, training happens in three stages.

1. Reading. The model is fed a vast amount of text: public web pages, books, articles, forums and other writing. It practises by guessing hidden words and correcting itself when it gets them wrong. This is repeated an astonishing number of times on powerful computers.
2. Guidance. A model that has only read can continue text, but cannot yet hold a polite conversation. So people write example conversations and rate its answers: this was good, this was poor, this was harmful. The model learns to be helpful and to decline harmful requests.
3. Polishing. Finally the model is tested and adjusted to make it as useful, honest and safe as possible.

One important consequence: training stops at a certain date. The model does not automatically know what has happened since. Some services can now search the web, but the underlying model is like an encyclopaedia printed on a particular day.

Why does AI sound so confident?

This may be the single most important thing to learn about AI. A language model is trained to produce fluent, believable text. But it does not always "know" whether something is true. When it does not remember something, it will not necessarily say "I don't know" – it may produce an answer that sounds right.

So an AI can invent, for example, a law that does not exist, the wrong opening hours or a book that was never written – and say it just as confidently as a real fact. This is called a hallucination (an error the AI has made up). Newer models do it less often, but it has not gone away.

In people, confidence often signals knowledge. In an AI, the tone tells you nothing about whether the information is correct. Read more in what AI can't do.

Does AI understand what it says?

Even researchers disagree on this. The honest answer: AI handles language in a way that looks like understanding, and it can reason surprisingly well. But it has no experiences, feelings or will of its own. It has never tasted coffee, never sat in a sauna and does not worry about its bills.

In practice this means AI is excellent with language and text, but its "knowledge" of the world is always second-hand – a summary of what people have written.

Does the AI remember me?

Within a single conversation, the AI remembers what you have discussed. That is why you can say "make it shorter" and it knows what you mean. When you start a new conversation it usually does not remember the previous one, unless the service has a memory feature switched on. In many services you can turn memory on or off in the settings.

The model itself does not learn from you in real time. However, the provider may use conversations to develop future models unless you have opted out in the settings. So it is worth checking your privacy settings, as we explain in security and privacy.

Try it: ask the AI to explain itself

A good way to understand AI is to ask it directly. Copy this into a chat AI:

Compare its reply with this guide. You will see the AI can be quite open about its limits – when you ask.

What to remember in everyday use

• Ask clearly. Because AI continues text, the better you describe your situation, the better the continuation. Read more in how to ask AI well.
• Check the facts. Fluent writing is not proof.
• Use it for drafts. AI is at its best when it writes the first version and you finish it off.
• Ask follow-up questions. You can always ask for reasons, sources or a simpler explanation.

Summary

A chat AI is a language model that predicts text one word at a time. It has learned from a huge amount of writing and from human feedback. That is why it writes fluently and knows a lot – but it can also be wrong in the same confident voice it uses when it is right. Once you understand this, you can use AI wisely: make the most of its strengths and check its weak spots.