The lessons › Lesson 1 of 6
What does a chat assistant actually do?
Why a chat assistant can sound sure of itself and still be wrong.
A chat assistant is built on a large language model: a program trained on a vast amount of text to predict which words are likely to come next. When you ask it something, it writes the answer a piece at a time, each time choosing a likely continuation. That is why it is fluent and quick, and good at drafting, summarising and rewording.
Predicting likely words is not the same as looking something up. Unless the tool is connected to a search engine or to documents you give it, it answers from patterns learned in training, and that training stops at a date. Where the pattern is strong but the fact is missing, it can produce something that reads well and is wrong. This is often called hallucination.
So treat an answer as a first draft from a well-read helper who rarely says "I don't know" unless asked to: useful and fast, and in need of checking whenever it matters.
An example
You ask for three books on the history of Belfast, with authors and years. You get three neat entries. Two are real. The third has a plausible title and a real historian's name, but no such book exists. Nothing in the answer marks the difference. (This is an illustration of a common pattern, not a record of a particular tool.)
Try this
Ask an assistant something you already know well: your town, your work, a hobby. Read the answer line by line and mark anything wrong, vague or invented.
Check yourself
Does a chat assistant look facts up when it answers?
Not unless it is connected to search or to documents you give it. Otherwise it writes from patterns learned in training.
Why can a wrong answer sound confident?
Because the model is choosing likely-sounding words. The fluency comes from the pattern, not from having checked the fact.
What AI cannot do here
It cannot tell you, by itself, which of its sentences are reliable. Some tools show sources; check that they exist and that they say what is claimed.
Sources
Last reviewed 1 October 2026. Glossary · Report a mistake