In simple terms
A large language model is a machine learning system trained on enormous amounts of text. Its core job is narrow: given some text, work out what is most likely to come next.
Do that over and over and the model writes sentences, answers questions or drafts code.
How it works
Text is first chopped into tokens, chunks roughly the size of a word or part of one. The model cannot handle words directly, because it works by adding and multiplying numbers, so every token is turned into a list of numbers.
The numbers inside the model are its parameters. During pretraining the model guesses the next token, checks the guess against real text, and updates those parameters. Guess, check, update, at enormous scale.
A model fresh out of pretraining is not yet what most people picture. It does not reliably follow instructions, and it mirrors the text it learned from, hate speech and dangerous material included. Later training stages are added to shape how it behaves.
Because the model produces what is likely rather than what is true, it can state inventions with complete confidence. That failure has a name: hallucination.
Why it matters
Language models now sit behind chat assistants, search summaries, translation, coding tools and customer support, often without being visible.
The confidence problem bites hardest in high stakes work. Stanford research found legal models hallucinating in roughly one in six benchmarked queries or more, while a system anchored to retrieved source documents reached 97.3 percent factual accuracy in simulated conversations.
Governments are legislating at very different speeds, so rules on disclosure, copyright and testing vary widely between markets for the same model.
Where you’ll see it
- Chat assistants and writing tools.
- Summaries at the top of search results.
- Coding assistants that complete or explain code.
- Customer support bots and machine translation.
Example
Asked for the capital of Australia, the model looks nothing up. It predicts the most likely continuation of the sentence, token by token, and produces Canberra because that is the pattern its training text supports.
Often confused with
A large language model is not a search engine or a database. It holds patterns rather than records, which is why it can be fluent and wrong at the same time.
Key facts
- Large language models work by taking an input and calculating what is most likely to come next, a skill acquired during pretraining.1
- Text is broken into tokens and converted into numbers, because the models add and multiply numbers and cannot process words directly.1
- The numbers inside a model are called parameters, and training adjusts them through repeated guessing, checking and updating.1
- A model straight out of pretraining does not reliably follow instructions and reproduces the text it was trained on, including hate speech and dangerous information.1
- Hallucination is when a system generates information that is incorrect, misleading or fabricated but presented as factual, and Stanford research found legal models hallucinating in about one in six benchmarked queries or more.2
Related concepts
In the news
Quick checkWhy can a large language model sound confident and still be wrong?Show answer
Because it predicts the most likely next token rather than checking a fact. Fluent text and true text are not the same target, and the gap is called hallucination.
Sources
- Center for Security and Emerging Technology, Georgetown University. The Surprising Power of Next Word Prediction: Large Language Models Explained, Part 1. 8 March 2024 (accessed 18 September 2026)
- Stanford Institute for Human-Centered Artificial Intelligence. What are hallucinations (in AI)?. Undated (accessed 18 September 2026)
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed September 21, 2026Researched and drafted with AI assistanceReport an issue