What is an LLM. The model predicts the next piece of text, it does not check facts
What a large language model is, how it is trained, why it makes things up with full confidence, and how a model differs from an app like Claude or ChatGPT.
In this article
In short
An LLM, a large language model, is a program trained on a huge amount of text that writes its answer piece by piece by predicting what fits next. That is why it can sound sure and still make things up, which Anthropic calls hallucination. Claude and ChatGPT are apps in which such a model works together with tools.
Someone in the office asks a chat about a rule in a new law and gets an answer with a section number, a title and a short quote. It sounds sure. The section does not exist. To see where that comes from, you need to know what an LLM, a large language model, is and what exactly it does when it answers.
A short definition is in our AI glossary. Here we go deeper. This article is part of our Claude Code guide, and we base the facts about models on Anthropic's glossary, as of 10 November 2026.
What is an LLM in a few sentences
LLM stands for large language model. Anthropic describes it as a language model with a very large number of parameters, trained on vast amounts of text, that can write, answer questions and summarise. Parameters are numbers inside the model that get set during training.
A model is not a database of ready answers. It has no table of laws and no catalogue of web pages to look things up in. It is a set of learned patterns between pieces of text. Both its strength and its weakness come from that.
How a language model is trained
Training has two stages. The first is pretraining. The model reads a huge collection of text and learns to predict the next word from what came before. Anthropic says plainly that a model after pretraining alone is not good at answering questions or following instructions.
The second stage is fine-tuning. The model is trained further on extra data and then with RLHF, reinforcement learning from human feedback. People compare several answers and pick the better one, and the model learns to write like the higher ranked ones. Only after this stage does it behave like an assistant that answers questions and follows instructions. Claude went through both stages.
Training also sets a knowledge cutoff. The model knows the world up to a certain point and has read nothing after it. Anthropic lists that date for each model in its models overview. For the current Claude models it is June 2026. The model knows about anything later only if it has web search or you paste it in.
Predicting the next token
The model does not see letters or whole words. Text is split into tokens, which are pieces of words, sometimes short whole words or single characters. We cover them in the article on tokens and the context window.
The answer is built token by token. The model looks at the whole conversation, works out which pieces fit next, picks one, adds it and does the same for the next one. Newer models can think before they answer, but that thinking is also text generated token by token and counts like the rest of the answer.
The choice of the next token does not always come out the same. A setting called temperature controls this. Higher temperature gives more varied answers, lower gives more predictable ones. Anthropic notes that even at temperature zero the same question can produce different answers. That is why the same report done twice can show different numbers, which we cover in the article on consistent results from AI.
Why an LLM makes things up
The model learned to write text that fits the context and resembles what it has read. From the description of training you can see that the goal was to predict the continuation well, not to check what is true. When you ask about something the model does not know well, it will still write a sentence that sounds right. The section number from the start of this article looks real because real section numbers look like that.
Anthropic calls this hallucination, and its guide to reducing hallucinations gives three basic methods. The first is to tell the model outright that it may say "I don't know". The second is for long documents. The model first writes out word for word quotes and only then works from them. The third is to make it give a source for every claim and drop anything it cannot back with a quote. The documentation adds that these methods reduce hallucinations but do not remove them, so important information still needs checking.
In an ordinary chat it can look like this.
Answer only from the attached contract. Add a word for word quote from the contract to every answer. If something is not in the contract, write "not in the document".
What an LLM is good at and what it is not
A model handles text you already have well. It summarises, rewrites in another tone, translates, turns notes into a table and writes the first draft of an email or a description. Current Claude models also read images and work in many languages.
It is weaker where a fact from outside the conversation matters. Dates, section numbers, prices, contact details. An answer from the model's memory is only a starting point to check against the source. With large tables, we have the model write a formula or a script and let the spreadsheet or the code do the counting. Then you can check the result instead of trusting it.
A good task description helps too. Context, goal and what the model must not do are covered in the article on how to write prompts.
Model vs app, or Claude and ChatGPT
Claude and ChatGPT are app names, and inside each one runs one or several models. The app adds the rest to the model. The chat window, memory, web search, files and links to email or a calendar. The Claude app has a model picker, and in Claude Code you change the model with the /model command.
Anthropic now has several models with different jobs. The strongest are for demanding reasoning and long agentic work, the lightest for fast tasks in large volume, such as classification or extracting data. How to match a model to a task is covered in which AI model to use.
The model on its own only writes text. When it gets tools, for example to read files or run commands, it can take the next steps by itself. Then we talk about an agent, which we describe in what is an AI agent.
Treat a model's answer like a draft from someone who has read a lot but checked nothing. Facts that go to a client or an authority should be confirmed at the source.
If you want to know where AI can take work off your company, ask for our free AI audit. We look at your tasks and tell you where to start.
What would you like to change in your business?
Leave your number. We will talk about your process and see whether we can help. The first call is free.
The service this text is about
Automation and AI / OCRAutomation of document re-typing, OCR and checking rules, and your data stays with you.