A large language model (LLM) is an AI model trained on vast amounts of text to predict what comes next, which lets it write, summarise, translate and answer questions in ordinary language. ChatGPT, Claude and Gemini are all built on large language models.
How does a large language model work?
Given some text, the model predicts the most likely next token (a chunk of a word), adds it, and predicts again. Repeated thousands of times, that produces a sentence, then a page, then a report. The predictive text on your phone does something similar with the last few words; a large language model takes in far more of what's in front of it and has learned from an enormous amount of writing.
The design comes from the transformer architecture introduced in the 2017 paper Attention Is All You Need. "Large" refers to the number of parameters and the amount of training data, not to the length of the answers.
Does an LLM look things up?
The model on its own doesn't. It generates text from patterns learned in training, and what it knows stops at a training cut-off date.
Many chat apps add a step around the model, though. They run a web search or pull passages from your documents, then hand the results to the model to work from. Claude, for example, can search the web and cite its sources in an answer (Anthropic, 2025). When an app draws on a set of documents this way, it's called retrieval-augmented generation. If an answer arrives with no sources attached, assume it came from the model's training alone.
Why can it sound right and still be wrong?
An LLM is trained to produce plausible text, and plausible text isn't always true. Fluent prose about a regulation, a date or a figure tells you nothing about whether it's accurate, and a confident wrong answer is called an AI hallucination.
The same method is what makes it useful. It can draft something sensible about almost any subject in seconds, including your own work, because it has seen a great deal of writing that resembles yours. The habit that makes it safe to rely on is asking "how would I check this?" before using anything that matters.
Suppose you paste one clause from a supplier contract and ask what it means. The model explains it using patterns from the many similar contracts it has seen. That's usually helpful. Occasionally it'll mention a 30-day termination notice that isn't in your text, because clauses like yours often contain one. Treat its explanation as a first reading and check it against the document.
LLM vs AI model
An LLM is one kind of AI model. Many newer LLMs also accept images and audio, so the line between a "language model" and a "multimodal model" has blurred, and in everyday conversation the terms are used interchangeably.
Common questions
Why do different models give different answers?
They're trained on different data, at different sizes, with different settings and instructions. The same question can get two different answers from two models, and both can be reasonable. Even the same model can word its answer differently each time you ask.
Is a large language model the same as ChatGPT?
No. ChatGPT is a product built on top of OpenAI's models. The model generates the text, while the product supplies the interface, the memory, the tools and the rules around it.
Does an LLM learn from what I type?
Not while you're talking to it, because the model is fixed once training ends. A provider may use your conversations to train future models, depending on your plan and privacy settings, and an app may save notes as memory. Our guide on what happens to your data explains what to check.