What Is a Large Language Model? How AI Chatbots Work

Popular AI assistants, such as ChatGPT, Claude, Gemini and Copilot, run on a technology called a large language model, or LLM. You don’t need a technical background to understand how it works. Once you have the basic idea, you’ll find it easier to use these tools well and to know when to trust them.

Start with autocomplete

Your phone probably suggests the next word as you type. If you write “See you”, it might suggest “tomorrow” or “soon”. It learned those guesses from patterns in a lot of text.

A large language model works on a similar idea, at a much larger scale. When you ask it something, it looks at everything written so far, including your question and its own answer up to that point. Then it predicts what text should come next. It adds that piece, then predicts again, and keeps going until the answer is finished.

So when you ask it to “write a polite email to my landlord about a broken heater”, nothing gets looked up in a database of emails. The model builds the email a little at a time, choosing what usually comes next in a polite email like that.

The analogy has limits. Phone autocomplete usually looks at only a little recent text. A large language model can take in pages of text at once and connect ideas that are far apart. That’s why it can follow a long conversation, keep the details of your heater problem in mind, and match the tone you asked for.

What “large” and “language model” mean

A language model is any system that predicts text. “Large” refers to two things:

• The amount of text it learned from. Modern models are trained on huge collections of writing: books, websites, articles, code and more.
• The size of the model itself. Inside, a model holds billions of adjustable numbers called parameters. You can think of them as tiny dials. Each one is set during training, and together they store the patterns the model picked up.

Models also don’t read whole words the way you do. They break text into small chunks called tokens. A token might be a whole word, part of a word, or a punctuation mark. When people say a model “predicts the next word”, it’s really predicting the next token. The idea is the same.

How a model learns

Training usually happens in two broad stages.

Stage one: learning from huge amounts of text. The model is shown text with the next part hidden and asked to guess it. When it guesses wrong, its dials get nudged so it would guess a bit better next time. This repeats an enormous number of times. Along the way, the model picks up grammar, facts, writing styles, and how ideas tend to connect. Nobody types in rules like “Paris is the capital of France”. The model absorbs patterns like that because they show up again and again in what it reads.

Stage two: learning to be a helpful assistant. A model that has only finished stage one is good at continuing text, but it isn’t yet good at answering questions or following instructions. So developers train it further on examples of good conversations. People often rate its answers too, and the model is adjusted toward the kinds of responses people preferred. This is a big part of why assistants feel conversational and polite.

Once training ends, the model’s knowledge mostly stays fixed. It doesn’t automatically learn about events that happened afterward. The date its training data ends is called its knowledge cutoff. Some assistants can search the web or read files you share, which helps them work with newer information.

Why it sounds so confident

People often notice that AI sounds equally sure whether it’s right or wrong. There are a few reasons.

• It learned from confident writing. Most of the text it trained on is written in a clear, settled tone. Articles and textbooks rarely stop to say “I’m not sure.” So the model learned that tone.
• Smooth wording doesn’t mean a checked fact. The model is very good at producing text that reads naturally. A made-up detail can come out just as smoothly as a true one, because both fit the pattern of a good answer.
• It was rewarded for being helpful. In stage two, answers that seemed complete and useful were often preferred. That can push a model toward giving an answer when saying “I don’t know” would have been better.

When a model states something false as if it were true, people call it a hallucination. A common example: you ask for sources for a school project, and you get a book title that sounds real but doesn’t exist. Developers are working to reduce this, but the problem hasn’t gone away.

What these models are good at, and where they fall short

Often useful for:

• Drafting and rewriting text: emails, cover letters, menu descriptions for a small bakery
• Summarising long documents
• Explaining a topic in simpler words
• Brainstorming ideas and options
• Helping with code, spreadsheets and formulas

Common limits:

• Facts can be wrong. Check anything that matters, especially names, numbers, dates, quotes and sources.
• Knowledge can be out of date unless the assistant is searching the web.
• Math and counting can slip. Models have improved a lot here, but double-check calculations.
• Each model knows only what it was trained on, what you give it, and what it can look up. It has no access to your situation unless you share the details. Avoid pasting passwords, account numbers, medical records or other sensitive personal information, and check the tool’s privacy settings.
• They can reflect biases in the text they learned from.

What’s settled and what’s still debated

Well established: LLMs are trained to predict text. They learn from very large amounts of writing. Most LLMs produce answers one piece at a time. They can be fluent and wrong in the same sentence.

Still debated: Whether these models “understand” anything in the way people do. Some researchers argue that predicting text this well requires building some kind of internal picture of the world. Others say it’s sophisticated pattern-matching and shouldn’t be called understanding. There are open questions about how well models really reason, and about what exactly happens inside them. Researchers can’t yet fully explain why a model produces a particular answer.

You don’t have to settle that debate to use these tools well. A practical way to think about an LLM is as a very well-read assistant that writes quickly and sounds sure of itself. Use it to get started faster, and check its work before you rely on it.