How do AI chatbots work?

What a chatbot is

A chatbot is a program you talk to by typing (or speaking), and it answers in ordinary language. Older chatbots, such as the help box on a bank website, followed a script: if you typed "balance", it showed a fixed reply. Modern AI chatbots are very different. They can answer almost any question, write a poem, explain a letter from the council or help plan a meal.

Behind them sits something called a large language model, often shortened to LLM. "Large" because it is enormous, "language" because it works with words, and "model" because it is a learned set of patterns (see What is AI, simply?).

The giant game of "what comes next?"

At its heart, a language model plays one game over and over: given some text, guess what comes next. If you see "Once upon a", you would probably guess "time". You learned that from reading lots of stories.

A language model learned the same kind of thing, but from a vast amount of text: books, articles, websites and more. From all that reading it picked up grammar, facts, writing styles and how ideas usually connect. It does not store the text like a library. It stores patterns, a bit like how you remember how a story usually goes rather than every word of it.

The model works with tokens, which are small chunks of text: a short word, part of a longer word, or a punctuation mark. When it writes a reply, it predicts one token, adds it to the text, then predicts the next, and the next, very fast. That is why you often see the answer appear word by word.

From text predictor to helpful assistant

A model that only predicts text would ramble. So companies add extra training to make it behave like a helpful assistant. People write example conversations and rate the model's answers as good or bad, and the model is adjusted to give more of the good kind. This teaches it to follow instructions, stay polite and refuse harmful requests.

Many chatbots can now also search the web, read a document you upload, or look at a photo. When they search, they read a few web pages and summarise them for you, which helps with recent news the model was not trained on.

Tip: The chatbot only "remembers" your current conversation (and any memory feature you have switched on). Start a new chat when you change topic for cleaner answers.

Why chatbots sometimes get things wrong

Because the model is choosing likely words, not checking facts, it can produce something that sounds right but is not. It might invent a quote, a statistic, a website link or a law. This is often called a hallucination.

Other common problems:

• Out-of-date knowledge. The model learned from text up to a certain date. Without a web search it may not know recent events.
• Agreeing too easily. If you push back, it may change its answer even when it was right the first time.
• Maths and counting slips. Some chatbots handle numbers well now, but small errors still happen.
• Missing context. It does not know your situation unless you explain it.

Getting better answers

What you type is called a prompt. Clear prompts get better replies. A few habits help a lot:

1. Say who it is for. "Explain this to a ten-year-old" or "Write this for my manager" changes the tone.
2. Give the details. Paste the text you want summarised, or list the ingredients you actually have.
3. Ask for a format. "Give me five bullet points" or "Keep it under 100 words".
4. Ask follow-up questions. "Make it friendlier" or "What did you leave out?" usually improves the result.
5. Ask for sources on factual questions, then open them yourself to check.

Try this

Ask a chatbot the same question twice in two new chats. You will often get slightly different wording, because it picks from several likely words each time. Then ask it something with a definite answer you can check, such as the opening hours of your local library, and compare with the library's own website. Before sharing anything personal, read using AI tools safely.