What is a large language model? A plain-English guide

A large language model (LLM) is a program that has learned the patterns of written language well enough to continue almost any text you give it. ChatGPT, Claude, Gemini and Microsoft Copilot are all built on one. Under the chat window they do the same thing: predict what comes next, one small chunk at a time. That explains both their strengths and their mistakes.

Tokens: how a model reads

A model doesn’t read words the way you do. It splits text into tokens: small chunks that might be a whole word, part of a longer word or a punctuation mark. OpenAI’s rule of thumb for English is that 100 tokens is roughly 75 words. Tokens are how limits are measured: how much you can paste in, and how long an answer can be.

Training and using are different things

Training happens before you ever see the model. It’s shown a vast amount of text and keeps adjusting a huge number of internal settings until it gets good at predicting the next word. Anthropic’s glossary calls this pretraining. Further training, including feedback from people on which answers are better, turns that text predictor into an assistant that follows instructions.

Training stops at a point in time: providers publish a knowledge cut-off for each model. After that date its knowledge runs out, unless the tool can search the web or read documents you give it.

Using it is different. Typing a question doesn’t change its settings, so it isn’t learning from you as you go: it works from its training plus whatever is in the conversation. Whether your chats train future models is a separate question (see below).

Why it predicts text, and why that matters

The UK government’s AI Playbook puts it plainly: the model weighs up possible responses “based on what it has learned from its training data” and presents the likeliest fit for your prompt.

Likely isn’t the same as true. When the right answer was common in the training data, the two usually line up. When it wasn’t (a niche regulation, a local detail, anything about your own organisation), the model still writes something fluent and confident. That’s a hallucination: in the playbook’s words, “content that appears plausible but may actually be factually incorrect”. We cover the causes and fixes in why AI gets things wrong.

What it’s good at, and what it isn’t

It’s good at:

• First drafts: letters, emails, reports, job adverts, funding bids.
• Summaries: long documents, meeting notes, email threads.
• Rewriting: plainer, shorter, more formal, or for a different reader.
• Sorting and extracting: pulling names, dates and amounts out of documents into a table.

It’s less reliable on:

• Facts it wasn’t given, especially recent, local or specialist ones.
• Details where a small slip matters: figures, dates, legal references.
• Judgement that needs context it can’t see: the phone call, the client’s history, last week’s agreement.
• Professional advice. As the playbook says, LLMs “are not a substitute for professional advice”.

In short, it does best when the material it needs is in front of it and a person can check the result quickly.

Context windows: how much it can see at once

The context window is everything the model can look at while it writes a reply: your instructions, any documents you’ve attached, the conversation so far and the reply itself. Anthropic describes it as the model’s “working memory”, separate from what it learned in training. Every tool has a limit.

Two practical consequences:

• If it isn’t in the window, the model can only guess. Questions about your own policies, prices or procedures need the right documents in front of it. Tools that look up the relevant passages first (called retrieval) are built for this: see how retrieval works.
• More isn’t always better. Anthropic’s documentation notes that as the amount of text grows, “accuracy and recall degrade”. Give it the documents that matter rather than everything you have, and start a fresh conversation when you change task.

Why the same question gets different answers

At each step the model has several possible next tokens, each with a likelihood, and it doesn’t always take the most likely one. A setting called temperature controls how much randomness goes into the choice: higher gives more varied answers, lower gives more predictable ones. Small changes in your wording, or in the conversation before it, also change what comes next.

That helps with brainstorming and hinders anything that must be consistent. For repeatable work, use the same written instructions every time, give it the source material, and have a person check a sample of the results. If something must come out exactly the same every time, such as a calculation or a standard clause, use a template or ordinary software instead.

Privacy: business plans and consumer plans

This is the part to get right before anyone pastes in a client’s email.

Consumer versions of these tools (the free and personal paid plans) may use what you type to train future models. OpenAI says that for its services for individuals, such as ChatGPT, “we may use your content to train our models”, with a setting to opt out. Since August 2025, Anthropic has asked users of its consumer Claude plans to choose whether their chats can be used.

Business plans are different. OpenAI says that for ChatGPT Business, Enterprise and its API (the route developers use to build it into other software), “By default, we do not use your business data for training our models.” Anthropic says much the same for its commercial products, such as Claude for Work and its API.

So, as a rule:

• For work, use a business plan your organisation controls, not personal accounts.
• Keep personal data and confidential client information out of consumer tools.
• Read the terms: where data is stored, how long it’s kept, and whether it’s used for training.
• UK GDPR still applies. Personal data needs a lawful basis, meaning a legal reason to use it, and the ICO (the UK’s data protection regulator) says that “in the vast majority of cases” using AI involves processing likely to result in a high risk, which means a data protection impact assessment (DPIA). If you can’t tell whether a tool is compliant, the ICO’s advice is to choose a different one.

For more, see our UK GDPR checklist for AI tools and choosing an AI assistant for your team. Or tell us about a real task.