What is an LLM? An LLM, short for large language model, is an AI model that generates text by repeatedly predicting which piece of text is likely to come next. An LLM learns this through training on vast amounts of text and is then refined into an assistant that answers questions.
An LLM is not the same as ChatGPT or Claude: the model is the engine, the app is the product built around it. There are several model families, from closed models by OpenAI, Anthropic and Google to open models you can run yourself.
An LLM has clear limits. It does not guarantee factual accuracy, it only has text up to its training date and it does not always give the same answer to the same question.
What happens to your input depends on the account: with personal accounts, the provider may use your text for training; with business accounts, it does not by default. That choice partly determines how you use an LLM in your work.
What does LLM stand for?
LLM stands for large language model: an AI model that generates text by predicting what is likely to follow, based on patterns from a vast amount of training text.
“Large” refers to two things: the number of parameters (the adjustable values in the model, which run into the billions) and the amount of text the model was trained on. An LLM is a form of generative AI focused on language, whereas other generative models create images, audio or video.
How does an LLM work?
An LLM works by breaking text into small pieces and choosing the most likely next piece each time. It learns what it needs for this in three steps.
How is an LLM trained?
An LLM is trained by having it predict the next piece of existing texts over and over, and adjusting the model when the prediction is wrong. After billions of these steps, the model has captured patterns in language: grammar, style, links between concepts and a lot of factual information.
That information is a collection of patterns, not a database of facts. The model also only has text up to the date its training stopped.
What are tokens in an LLM?
Tokens are the pieces of text an LLM computes with: a whole word or part of one. The model predicts what follows token by token, each time using all the preceding text as context.
The architecture that makes this possible is the transformer, described by Google researchers in 2017. That same transformer is the T in GPT, as you will see when you look up what GPT stands for.
How does an LLM become an assistant?
An LLM becomes an assistant through additional training on instructions and on ratings by people. Without that phase, a model only completes texts and does not answer questions. That second phase explains much of how ChatGPT works in a conversation.
What is the difference between an LLM and ChatGPT?
The difference between an LLM and ChatGPT is that the LLM is the model and ChatGPT is the product around it. The product adds an interface, a hidden system instruction, memory between conversations and tools such as web search or reading files.
- Interface
- System instruction
- Memory
- Tools such as search and files
The same model can answer differently in the app and through the API.
That is why the same model can answer differently in the app and through the API, the connection developers use to call the model directly. What works in claude.ai does not always work identically for me through the API, because the system prompt and the interface context are missing there. If you want to understand ChatGPT as a product, you need to keep both layers apart.
Which LLMs are there?
There are six LLM families that are most relevant to professionals in Europe:
- GPT from OpenAI, the model behind ChatGPT.
- Claude from Anthropic, available through claude.ai and the API.
- Gemini from Google, built into Google Workspace and the Gemini app.
- Llama from Meta, with publicly shared model weights; since April 2026, Meta has also released the Muse models, of which Muse Glimmer and, since 15 October 2026, Muse Spark Open are open-weight; Muse Spark itself remains closed.
- Mistral from the French company Mistral AI, partly open and partly closed.
- DeepSeek from the Chinese company of the same name, with public model weights.
The main distinction between these families is open versus closed. You use a closed model through its maker’s service; you can run an open model on your own servers, so your text does not leave the organisation. The trade-off between open source AI and closed models is mainly about control over your data versus convenience and quality.
What can’t an LLM do?
There are four things an LLM cannot do that users often expect from it:
- Guarantee factual accuracy: the model generates what sounds likely, even when it is wrong.
- Provide current information: without a search feature, the model has nothing on events after its training date.
- Calculate reliably: calculations go wrong more often unless the product calls in a calculation tool.
- Check sources: a model can cite a source that does not exist, and does not notice this itself.
The first and last points together are called AI hallucination. During the period I worked intensively with ChatGPT, I saw that pattern recur: studies that turned out not to exist when I checked, and search volumes for keywords even though the model had no connection to any keyword tool.
Why does an LLM give a different answer each time?
An LLM gives a different answer each time because, for every token, it chooses from several likely options with an adjustable degree of randomness. Small differences at the start of an answer grow into different wording, and sometimes a different conclusion.
Even with randomness set low, the Claude API does not always give identical output for the same input in my own scripts. Since API version 2026-10-21, you can no longer set that randomness in the Claude API, because temperature, top_p and top_k have been removed there. In practice, this means you should not expect one right answer: use a fixed prompt, possibly with examples through few-shot prompting, and check the result, especially for tasks that must give the same outcome every time.
What happens to the text you enter into an LLM?
Text you enter into an LLM is processed on the provider’s servers, and with personal accounts the provider may use that text to train future models. For ChatGPT personal accounts, this is on by default and you can turn it off in the settings; for Claude’s consumer plans, it depends on your setting.
For business accounts and the API, OpenAI and Anthropic do not use your input for training by default. So never put client data or confidential documents into a personal account.
Frequently asked questions about LLMs
You also come across the abbreviation LLM outside AI, and a few technical questions surround the concept itself.
Is an LLM the same as an LL.M.?
No, an LLM in AI is a large language model, whereas LL.M. stands for Master of Laws, a master’s degree in law. In the Netherlands, according to Studiekeuze123, you receive that title after a master’s in law at a university or university of applied sciences, and you may put the abbreviation after your name. After a university master’s in law, you may also use the Dutch title mr.
What is an LLM’s context window?
An LLM’s context window is all the text the model can take into account for one answer, including the answer itself. Anthropic describes it as the model’s working memory, separate from the data it was trained on earlier. The window is measured in tokens, and if your input is larger, the Claude API rejects the request. According to Anthropic, accuracy can also decline as the window fills up.
Can an LLM also process images?
Yes, many current LLMs also process images; a model like that is called multimodal. Google writes that the Gemini models were built to be multimodal from the start, and the Gemini API accepts images and video as well as text. Through the API, you can send an image and ask a question about it, without needing a separately trained model.
What does an LLM mean for your work?
For your work, an LLM is a fast writer and an unreliable source of facts at the same time. An LLM produces a first draft, a summary or a structure in seconds, but the judgement on what is correct stays with you. Once you know that distinction, you can use an LLM where it performs well and check the rest.
LearnLLM’s AI Literacy Course covers exactly that basis: how a large language model produces text, where it goes wrong and how to use an LLM in a way that lets you account for the output.
