An LLM (large language model) is a computer program that has read an enormous amount of text and learned to predict which word comes next. That simple skill, repeated very fast, lets it answer questions, write emails, summarise documents and translate. ChatGPT, Claude and Gemini are all built on LLMs. An LLM does not look facts up like a search engine: it writes the most likely answer, which is usually helpful and sometimes wrong.
An everyday analogy
Think of the autocomplete on your phone. You type "See you" and it suggests "tomorrow". It has seen that pattern many times.
An LLM is the same idea scaled up enormously. Instead of suggesting one word from your own messages, it has learned patterns from a vast library of books, websites and other text. So instead of finishing "See you", it can finish "Write a polite letter to a patient explaining that their appointment has moved", one word at a time, until the letter is done.
What the words mean
- Large: the model is trained on a huge amount of text and has billions of internal settings (called parameters) that store what it learned.
- Language: it works with words. Many newer models also handle images and audio, which we explain in what is multimodal AI.
- Model: a mathematical system that turns an input (your question) into an output (its answer).
LLMs do not read text as whole words. They split it into small pieces called tokens. Our guide what are tokens in AI explains why that matters for cost and length.
How an LLM learns
Training happens in stages, long before you type anything:
- Reading: the model is shown a vast amount of text and practises guessing the next piece, over and over. Each wrong guess slightly adjusts its parameters. This is called pre-training.
- Learning to be an assistant: people write example questions and good answers, and the model is trained on those so it behaves like a helpful assistant rather than a text predictor.
- Feedback: people compare different answers and rank them. The model is tuned towards the kind of answers people prefer.
Once trained, the model is fixed. When you chat with it, it uses what it learned, plus whatever you put in front of it in the conversation. For the step-by-step version, see how does ChatGPT work.
A concrete example (fictional)
A small law firm, "Hartley & Moss", receives a 40-page lease. A paralegal asks an LLM: "List every clause in this lease that mentions early termination, with the clause number."
The LLM reads the lease (it is part of the conversation now) and writes a list. It does this well because it has seen countless contracts during training and knows what termination clauses tend to look like.
But the paralegal still checks the clause numbers against the original. Why? Because an LLM writes the likely answer. If a clause is unusual, the model can miss it or, more rarely, invent a plausible-sounding clause number. This is called a hallucination.
What LLMs are good and bad at
Good at:
- Drafting and rewriting text (letters, summaries, reports)
- Explaining complex material in plain language
- Translating and changing tone
- Pulling structure out of messy notes
Weaker at:
- Exact facts it was never shown, or recent events after its training
- Precise arithmetic over long tables
- Knowing when it does not know
The fix is usually to give the model the source material (a document, a database, a search result) and ask it to work only from that. That is the idea behind retrieval, explained in our RAG guide.
Different LLMs, different strengths
There is no single LLM. OpenAI makes the GPT models, Anthropic makes Claude, Google makes Gemini, and there are open-source models too. They differ in writing style, reasoning, speed and price. Many people find one model better for drafting and another for analysis. You can compare them on our models page.
What this means for confidential data
Practical tip: before you paste a patient letter or client file into any LLM tool, check the provider's terms for three things: whether your input is used for training, how long it is kept, and whether the provider takes a licence to reuse it.
That last point matters more than most people think. Data protection law, such as the GDPR, protects information about identifiable people. If a provider may reuse your input, it can strip the names and the rest is no longer protected: it can legally be pooled, licensed and sold. We show how that works on the data trail.
Where Wysor fits
Wysor gives you several LLMs, including GPT, Claude and Gemini, in one workspace. We act only as your processor under Art. 28 GDPR, take no licence to reuse your content, and never train on it or sell it, anonymised or not. Data is encrypted, hosted in the EU, and every connected AI model provider runs with zero data retention: processed, then deleted.
FAQ
What does LLM stand for? LLM stands for large language model: a program trained on a very large amount of text to understand and generate language.
Is ChatGPT an LLM? ChatGPT is an app built on top of LLMs made by OpenAI. The app adds the chat window, history, file uploads and safety rules around the model.
Does an LLM think or understand? It produces language by recognising patterns, which can look like understanding. It has no beliefs or intentions, and it can state wrong things confidently, so important output should always be checked.
Does an LLM learn from my chats? Not while you chat: the model itself does not change during your conversation. But some providers may use conversations later to train future models, depending on your plan and settings. Check the terms.
Further reading
Questions about using AI with sensitive data? Reach us at [email protected] or through the in-app contact form.