A large language model is an artificial-intelligence system trained on enormous amounts of text to predict the most likely next piece of a sentence, and through that single skill it can generate, summarise, translate, and answer questions in language that reads as if a person wrote it.
Large language models power many of the AI writing and chat tools that became widely used in the mid-2020s. The name describes what they are: models of language that are large, both in the volume of text they learn from and in the number of internal settings, called parameters, they use to represent patterns. Understanding how they work — and where their limits lie — helps separate genuine capability from hype.
How does a large language model actually work?
At its core, a large language model does something surprisingly simple to describe: it predicts the next token given the text so far. Because computers work with numbers rather than words, text is first broken into tokens, which can be whole words, fragments of words, punctuation, or spaces. Each token is turned into a list of numbers called an embedding, and these numbers begin as essentially random values.
During training, the model reads vast quantities of text and repeatedly tries to guess the next token. Each time it guesses, it compares its prediction with the actual next token and nudges its internal numbers to do slightly better next time, a process repeated across billions of examples. Over time, the embeddings and other settings come to capture how words relate to one another — which words tend to follow which, how ideas connect, and how grammar and meaning are structured. No one hand-codes these rules; they emerge from the patterns in the data.
What is the transformer, and why does it matter?
The breakthrough that made modern large language models possible is an architecture called the transformer, introduced by researchers in the late 2010s. Its key innovation is an “attention” mechanism that lets the model weigh how relevant every other word in a passage is to the word it is currently processing. That allows it to keep track of context across long stretches of text — understanding, for example, that a pronoun near the end of a paragraph refers to a name near the beginning. Stacking many of these attention layers together, and training them at scale, is what turns simple next-word prediction into fluent, coherent output. Before the transformer, models struggled to hold on to context over longer passages, so this ability to look across an entire input at once was a decisive step forward for the field.
How is a large language model trained?
Training generally happens in stages. The table below outlines the phases in broad terms; exact methods vary between developers.
| Stage | What happens | Purpose |
|---|---|---|
| Pre-training | The model learns from a very large collection of text by predicting the next token | Builds broad language ability and general knowledge patterns |
| Fine-tuning | The model is further trained on more specific or higher-quality examples | Sharpens usefulness for particular tasks or styles |
| Alignment | Human feedback and guidelines steer responses | Makes outputs more helpful, safer, and better at following instructions |
The result is a system that has effectively absorbed statistical patterns from an enormous body of writing. It is worth stressing that a model does not store a tidy database of facts it can look up; it stores patterns, and it reconstructs plausible answers from them. Our artificial-intelligence section follows how these methods continue to evolve.
What can large language models do — and what can they not?
Within language, the range is broad: drafting and editing text, summarising documents, answering questions, translating, writing and explaining code, and holding a conversation. These strengths flow directly from having learned how language is structured across countless examples.
The limits flow from the same source. Because a model produces the most plausible continuation rather than verified truth, it can generate confident statements that are simply wrong — often called hallucinations. It has no built-in sense of whether a claim is accurate, no genuine understanding of the world in the human sense, and its knowledge reflects the text it was trained on, which has a cutoff and may be out of date. It can also reflect biases present in that training text. For a fuller treatment of these trade-offs, see our guide to what generative AI can and cannot do, and our broader technology coverage for how the field is developing.
Is a large language model the same as artificial general intelligence?
No. Despite impressive fluency, a large language model is a narrow tool specialised in language, not a general-purpose mind. It does not have goals, beliefs, or awareness, and it does not reason across every domain the way a person can. Whether scaling these models further leads toward more general intelligence is an active area of research and debate, but today they are best understood as powerful pattern-based language engines rather than thinking beings.
Why the word “large”?
The “large” refers to scale in two senses. First, the training data spans a huge and diverse collection of text. Second, the models contain a very large number of parameters — the adjustable internal values that store learned patterns — often numbering in the billions. In general, greater scale has tended to produce more capable and more general behaviour, which is a big part of why these systems improved so quickly. Scale also brings costs, including significant computing power for training and running the models, which is one reason efficiency has become a major focus of research.
Frequently asked questions
Do large language models actually understand what they say?
That is a genuine and unsettled debate. Large language models capture deep statistical patterns in language and can produce responses that appear to reflect understanding, but they do not have beliefs, intentions, or lived experience the way people do. They generate text by predicting likely word sequences, which is powerful yet different from human comprehension.
Why do large language models sometimes make things up?
Because they generate the most statistically plausible continuation rather than looking up verified facts, they can produce confident but incorrect statements, often called hallucinations. The output reads fluently even when the underlying claim is wrong, which is why important facts should be checked against reliable sources.
What is a token?
A token is the unit of text a model actually processes. It may be a whole word, part of a word, a punctuation mark, or a space. Text is broken into tokens, converted into numbers, and only then handled by the model, which is why usage and limits are often measured in tokens rather than words.
What is the difference between a large language model and a chatbot?
The large language model is the underlying engine that predicts and generates text. A chatbot is an application built around such a model, adding a conversational interface, safety measures, and sometimes access to tools or live data. One model can power many different products.
Are large language models the same as general artificial intelligence?
No. They are narrow systems specialised in language tasks, not general-purpose minds. They can be remarkably capable within language, but they do not possess broad, human-like reasoning across every domain, and researchers do not consider them to be general artificial intelligence.
