An AI chatbot that does not make things up
Language models produce fluent text whether or not they know the answer. Left alone on a website, one will tell a customer about a discount that does not exist. This is the main reason businesses distrust chatbots, and they are right to.
Why it happens
- A language model's default is to continue plausibly, not to verify.
- Without a defined source of truth, everything it says is a guess with good grammar.
The rule that stops it
Restrict the assistant to a named body of text, the business's own documents, and give it exactly one alternative to answering: saying the answer is not there and passing the question to a human.
- Every answer is traceable to an uploaded document.
- "I don't know" is a designed outcome, not a failure state.
- The human stays in the loop for precisely the questions that need one.
How to test any chatbot for this
- Ask something plausible that is not in the source material.
- An honest system declines and escalates. A dishonest one answers.
FAQ Plus is built on this rule. No guessing is the product.