A large language model predicts text, piece by piece, based on patterns learned from huge amounts of writing. It is strong at language tasks like drafting and summarizing, limited by what it was trained on and what you give it, and prone to confident mistakes. Knowing how it works helps you decide where to trust it.
When people talk about AI in business today, they usually mean tools built on large language models. ChatGPT, Claude, Gemini, and Copilot all run on this kind of technology. You do not need to understand the math to use them well, but a basic picture of how they work makes it much easier to know when to trust them.
The short version: a large language model is a very capable text predictor. Almost everything it does well, and almost everything it gets wrong, follows from that.
How a language model works
Google's machine learning course defines a language model as something that "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." A token is a small chunk of text, often part of a word. Google's course notes that one token is about four characters, or roughly three quarters of a word.
A large language model applies that idea at enormous scale. In Google's words, LLMs "predict a token or sequence of tokens, sometimes many paragraphs worth of predicted tokens." When you ask a question, the model generates the answer one token at a time, each time choosing a likely next piece based on everything before it.
Why they got so good so quickly
Most modern LLMs are built on an architecture called the Transformer, introduced in a 2017 research paper by Ashish Vaswani and colleagues titled "Attention Is All You Need." The key idea let models weigh how every part of a passage relates to every other part, rather than reading strictly word by word.
Combined with far more training data and computing power, that design produced models that can hold a coherent conversation, follow instructions, write in different styles, and work across many subjects.
What an LLM knows, and what it does not
An LLM learns patterns from the text it was trained on. It does not look things up by default, and it does not know about events after its training data was collected unless a tool gives it that information.
It also has a working memory for each conversation, called the context window. Anthropic's documentation describes the context window as all the text a model can reference when generating a response, and notes it is different from the large body of data the model was trained on. Anything you paste in, such as a contract or a price list, sits in that working memory for the conversation. It does not permanently teach the model.
What LLMs are good at
These are all language tasks. The model is working with words, and it is often very good at it.
Where they fail
Google's course states it directly: "LLMs hallucinate, meaning their predictions often contain mistakes." Because the model predicts likely text rather than checking facts, it can invent numbers, citations, or details that sound right. It can also be inconsistent, giving different answers to the same question.
LLMs are weaker at exact arithmetic, at knowing what they do not know, and at anything that depends on current or private information they have not been given. The practical rule is simple: use them freely for drafts and thinking, give them your real information when facts matter, and have a person check anything that will reach a customer or drive a decision.
When people talk about AI in business today, they usually mean tools built on large language models. ChatGPT, Claude, Gemini, and Copilot all run on this kind of technology. You do not need to understand the math to use them well, but a basic picture of how they work makes it much easier to know when to trust them.
The short version: a large language model is a very capable text predictor. Almost everything it does well, and almost everything it gets wrong, follows from that.
Google's machine learning course defines a language model as something that "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." A token is a small chunk of text, often part of a word. Google's course notes that one token is about four characters, or roughly three quarters of a word.
A large language model applies that idea at enormous scale. In Google's words, LLMs "predict a token or sequence of tokens, sometimes many paragraphs worth of predicted tokens." When you ask a question, the model generates the answer one token at a time, each time choosing a likely next piece based on everything before it.
Most modern LLMs are built on an architecture called the Transformer, introduced in a 2017 research paper by Ashish Vaswani and colleagues titled "Attention Is All You Need." The key idea let models weigh how every part of a passage relates to every other part, rather than reading strictly word by word.
Combined with far more training data and computing power, that design produced models that can hold a coherent conversation, follow instructions, write in different styles, and work across many subjects.
An LLM learns patterns from the text it was trained on. It does not look things up by default, and it does not know about events after its training data was collected unless a tool gives it that information.
It also has a working memory for each conversation, called the context window. Anthropic's documentation describes the context window as all the text a model can reference when generating a response, and notes it is different from the large body of data the model was trained on. Anything you paste in, such as a contract or a price list, sits in that working memory for the conversation. It does not permanently teach the model.
These are all language tasks. The model is working with words, and it is often very good at it.
Google's course states it directly: "LLMs hallucinate, meaning their predictions often contain mistakes." Because the model predicts likely text rather than checking facts, it can invent numbers, citations, or details that sound right. It can also be inconsistent, giving different answers to the same question.
LLMs are weaker at exact arithmetic, at knowing what they do not know, and at anything that depends on current or private information they have not been given. The practical rule is simple: use them freely for drafts and thinking, give them your real information when facts matter, and have a person check anything that will reach a customer or drive a decision.
ChatGPT is a product built on large language models made by OpenAI. Claude, Gemini, and Copilot are similar products built on their makers' models. The chat app adds features like memory, file uploads, and web search on top of the model.
It processes patterns in language extremely well, which can look like understanding. It does not check facts the way a person does, which is why it can be fluent and wrong at the same time.
Pasting text puts it in the model's working memory for that conversation. Whether a vendor later uses your data to train future models depends on the product and your settings, so check the vendor's data policy.
A token is a small chunk of text the model reads and writes, often part of a word. Google's course puts it at about four characters or three quarters of a word. Tokens also determine usage limits and, for many tools, cost.
It depends on the task, cost, data policy, and which tools your team already uses. Test two or three on your own real tasks and compare the results, rather than relying on general rankings.
This page simplifies how large language models work to make the ideas usable for business decisions. Products built on LLMs differ in features, data handling, and accuracy, and they change often. It does not evaluate or recommend any specific product.