Cost
A long chat sends the earlier turns again. More text usually means more for the provider to process. You do not need a price list to notice that pasting a whole document is a bigger request than pasting one paragraph.
Home · Start here · 01 · What a chatbot is doing
Lesson 1 of 7. Next: Prompts that hold up.
Lesson · No account
What a chatbot is doing
A chatbot reply is new text continued from the conversation so far. It is not a lookup in a database unless the product adds one.
1 What it is
The model writes the reply one piece at a time. Each piece is a likely continuation of the text so far, not a fact it has checked. NIST's Generative AI Profile says large language models predict the next token or word, and that this statistical prediction can be factually accurate or factually inaccurate.
A chat product can also search or call a tool. When it does not, the reply is only the continuation. Fluent writing is what that prediction is good at. It is not a sign that a source was opened.
Example
Question. What year did Northwind Kitchens open its Oslo shop?
Reply. Northwind Kitchens opened its Oslo shop in 2014.
This page invented the company and the year. Nothing was looked up.
2 Why it matters
Quality
A specific detail, such as a year or a name, can appear in a sentence that nothing in the chat supports. The tone does not tell you which sentences those are.
Risk
If you pass the reply on as if it were a source, the unchecked claim travels with your name on it.
3 How to apply · try today
Ask the chatbot you already use one question you can answer from a document you have. Compare the reply to that document. Mark one sentence that matches, and one that does not, if there is one.
Common mistake
Treating a specific detail as proof. A year, a citation, or a confident tone can all be part of the continuation.
Self-check
The chat did not search, and the reply names a year for a shop you cannot find. What is that year?
Pick a question you can check in a file or a page you already trust. Ask it in the chatbot you already use. Do not paste secrets. Then open the file and mark what matches and what does not.
Predicting the next piece of text is why a reply can sound finished when it is wrong. Anthropic's guide on reducing hallucinations and OpenAI's guide on accuracy both treat unsupported claims as something you design against, not as a rare glitch.
The Tokens card, linked below, is the builder version of what those pieces of text are.
Go further
The Tokens card is the builder version of this idea: the pieces a model reads, writes, and bills for.
Questions
- Does a chatbot look things up?
- Not by default. Some products add search or tools. If that did not happen, the reply is a continuation of the conversation, not a source you can cite.
- Why does a wrong answer sound sure?
- The model is producing likely words. Likely words are not a check against a document.