In short
A large language model (LLM) is a type of AI model trained on huge collections of text to predict what comes next in a sequence. ChatGPT (OpenAI), Gemini (Google) and Claude (Anthropic) are all built on LLMs. The model does not look facts up by default. It generates text that fits the patterns it learned, which is why it is fluent and also why it can be wrong.
How a language model works
Text is split into small pieces called tokens, which are whole words or parts of words. During training the model sees billions of examples and adjusts its internal settings until it becomes good at predicting the next token. After that, it is usually tuned further with human feedback so that it follows instructions and declines harmful requests.
When you type a question, the model writes its answer one token at a time, each choice based on everything before it. It has no memory of you between separate conversations unless the product it sits inside stores and reloads that history.
Why models get things wrong
The usual name for a confident wrong answer is a hallucination. It happens because the model is predicting plausible text. It has no built-in test for whether a statement is true. Names, figures, dates, quotes and legal or technical citations are where this shows up most.
Models also have a training cut-off. Unless the product connects the model to search or to your own documents, it knows nothing that happened after that date, and nothing about your business at all.
Context windows and your own data
Everything the model can consider in one go, your instructions, the documents you paste and its own reply, has to fit in its context window. Modern models have large windows, but they are still finite, and long inputs cost more to process.
For business work the reliable pattern is to give the model the facts it needs at the moment it answers: the relevant policy, price list or customer record. That technique is called retrieval-augmented generation, and it is covered in its own explainer.
What this means for a business
| Good at | Drafting, summarising, classifying enquiries, extracting details from messages, rewriting for tone. |
|---|---|
| Needs checks | Anything with figures, prices, legal wording or a promise to a customer. |
| Not suited to | Being the only source of truth, or making significant decisions about people with no human review. |
| Data question | Where the text you send is processed and stored, and whether it is used for training. Business plans and API terms usually differ from free consumer apps. |
When this matters
Useful background before any AI project. Every product New Era AI builds that writes or understands language, including Echo One, runs on language models with rules and checks around them.
Questions
Is a large language model the same as ChatGPT?
ChatGPT is a product. The large language model is the engine inside it. Many products use the same kinds of model with different instructions, tools and data around them.
Can we stop a model from making things up?
You can reduce it a great deal by giving the model the right facts at answer time, limiting what it is allowed to say, and checking outputs that matter. You cannot remove the risk entirely, which is why important outputs keep a human in the loop.
Does using an LLM mean our data trains the model?
It depends on the product and the plan. Many business and API agreements state that customer data is not used for training by default, while some free consumer tools may use conversations unless you opt out. Check the terms for the exact tool you use.