What Large Language Models Actually Do

The short version

Large language models — the technology behind ChatGPT, Claude, Gemini, and similar tools — are sophisticated text prediction engines. Given a sequence of words, they predict what comes next. That single mechanism, scaled to billions of parameters and trained on vast quantities of text, produces behaviour that looks remarkably like understanding.

But looking like understanding and actually understanding are different things, and the distinction matters enormously when you are making decisions about how to use these tools in your organisation.

How the prediction engine works

An LLM is trained by reading enormous amounts of text — books, articles, websites, code repositories — and learning statistical patterns about which words and phrases tend to follow which others. During training, the model adjusts millions (or billions) of internal parameters to get better at this prediction task.

When you type a prompt, the model generates a response one token at a time. Each token is chosen based on probability distributions shaped by the training data and the conversation so far. The result is fluent, coherent text that can answer questions, summarise documents, write code, and much more.

What this means in practice

The key implication is that LLMs do not have a database of facts they look up. They have patterns. This is why they can generate plausible-sounding statements that are entirely wrong — a phenomenon known as hallucination. It is not a bug to be patched; it is a natural consequence of how the technology works.

For professionals evaluating AI tools, this means that LLMs are best used for tasks where the output can be verified, where creativity and fluency matter more than factual precision, or where they are combined with retrieval systems that ground their responses in real data.