What Is a Large Language Model? A Plain-English Guide

If you’ve used ChatGPT to draft an email, asked Google’s Gemini to summarise a document, or noticed Microsoft Copilot suggesting text as you type, you’ve already met a large language model, or LLM. It’s the technology behind most of the “AI chat” tools that have appeared in the last few years, and it’s now built into search engines, office software, phone keyboards and customer service chatbots.

But what is it actually doing under the hood? Not magic, and not quite what most people assume. An LLM isn’t a database of facts you’re chatting with, and it isn’t a search engine either, even though it can feel that way when it gives you a confident, well-written answer. Understanding the difference matters, because it explains both why these tools are genuinely useful and why they sometimes get things wrong with total confidence.

This guide walks through what a large language model is, how it works in everyday language rather than technical jargon, where it shines, where it still trips up, and how the well-known tools like ChatGPT, Claude and Gemini fit into the wider picture. It’s a general explainer rather than a technical manual, so if you’re building something on top of a specific model for work or a project, it’s worth checking that provider’s own documentation too.

What is a large language model, in plain English?

A large language model is a piece of software trained on enormous quantities of text so that it can predict what word, or part of a word, is likely to come next in a sentence. That’s the core trick. Everything else, from writing essays to answering questions to producing code, is built on top of that one basic skill: guessing the most plausible next piece of text, over and over, until a full response has been built.

The technology behind almost all modern LLMs is called a “transformer”. Before transformers became standard, earlier language models tended to read text in strict order, word by word, which made it hard for them to keep track of context over long passages. Transformers instead weigh up how every word in a piece of text relates to every other word at once, which is part of why modern models are much better at holding onto context across a long conversation or a lengthy document.

During training, the model is shown vast volumes of text drawn from sources such as books, articles, websites and other written material, and it gradually adjusts itself to get better and better at predicting what comes next. This process requires huge amounts of computing power and, correspondingly, huge amounts of text.

The word “large” in large language model refers to that scale, both the amount of text used in training and the number of internal adjustable settings inside the model, known as parameters. Generally speaking, more parameters and more training data allow a model to capture more nuance in language, though scale isn’t the only thing that determines how good a model feels to use in practice.

How does a large language model actually work?

Before an LLM can do anything with your text, it has to break it down into small chunks called tokens. A token might be a whole short word, part of a longer word, or even just a punctuation mark. The model then uses patterns learned during training to predict the next token, one step at a time, feeding each new token back in as part of the context for the one after it.

So when you ask an LLM a question and watch the answer appear, what you’re actually seeing is the model repeatedly asking itself “given everything written so far, including my own answer up to this point, what’s most likely to come next?” It does this thousands of times per response, extremely quickly, and the result usually reads as fluent, coherent text.

Most publicly available LLMs go through an additional stage after their initial training, often involving human feedback. People review the model’s responses and rate them, and the model is adjusted to produce answers that are judged more helpful, better behaved, and less likely to be harmful, off-topic or nonsensical. This fine-tuning stage is a big part of why a modern chatbot feels noticeably more useful and better-mannered than an early, raw language model would.

It’s worth being clear about one thing this process does not involve: the model looking answers up in a fixed reference source. Unless it has been specifically connected to a search tool, a database, or some other external system, an LLM is generating language based on patterns, not retrieving verified facts. Some modern tools do combine an LLM with live web search or document lookup, which changes this picture somewhat, but the underlying language generation still works the same way.

Where do you actually encounter large language models?

LLMs sit behind a growing list of everyday tools, even when the branding doesn’t always make that obvious. Some of the best-known model families include:

• GPT, developed by OpenAI, which powers ChatGPT and is also used by some third-party products and services.
• Claude, developed by Anthropic, offered as a standalone chatbot and used in various business tools.
• Gemini, developed by Google, integrated into Google’s search, Workspace apps and Android devices as well as offered as a standalone assistant.
• Llama, developed by Meta, notable for being released with openly available model weights that other developers can build on.
• Copilot, Microsoft’s assistant built into Windows and Office, which uses OpenAI’s models under the hood.

These tools differ in tone, style, strengths and the way they’re integrated into other software, but they’re all fundamentally doing the same underlying job: predicting text based on patterns learned from training data.

Model family	Made by	Where you’ll typically meet it
GPT	OpenAI	ChatGPT and various third-party apps
Claude	Anthropic	Claude chatbot and some business software
Gemini	Google	Google Search, Workspace, Android
Llama	Meta	Open-weight, used by developers and other apps
Copilot	Microsoft	Windows, Office, Bing

Because most of these tools run in the cloud rather than on your own device, using them smoothly at home usually comes down to having a decent, stable internet connection rather than needing especially powerful hardware. If you’re wondering whether your home network is up to scratch for this kind of everyday use, it’s worth reading our guide to Wi-Fi 7 to understand whether upgrading is actually worth it for you.

What can a large language model do well?

LLMs tend to excel at tasks that are fundamentally about language, since that’s what they were trained on. In practice, that tends to cover things like:

• Drafting emails, messages or first versions of documents.
• Summarising long articles, reports or meeting notes into something shorter and clearer.
• Explaining a concept in simpler terms, or at a different reading level, for a different audience.
• Brainstorming options, headlines, names or different ways to phrase the same idea.
• Helping write or explain code, though the output still needs checking by someone who understands what it’s meant to do.
• Rewriting text in a different tone, for example making something more formal or more concise.

What ties all of these together is that they’re jobs about manipulating language itself, rather than jobs that require certainty about facts in the world. A good way to think about it is that an LLM is excellent at helping you say something, and much less reliable at telling you whether something is true.

If you’re using an LLM to help with everyday admin such as drafting replies to companies, filling in forms, or organising your accounts, it’s worth reading our guide on choosing a password manager, since you should never paste real passwords, PINs, or full account details into a chat tool. Treat these tools the same way you’d treat a helpful but very chatty colleague who has no obligation of confidentiality.

Where do large language models still get things wrong?

The best-known weakness of an LLM is “hallucination”, where the model states something confidently that simply isn’t true. It might invent a fact, misquote a source, get a date wrong, or fabricate a reference that looks entirely plausible but doesn’t actually exist. This happens because the model is generating the most plausible-sounding continuation of your