Glossary
Hallucination
A hallucination is a statement by a language model that sounds convincing but is factually wrong or entirely made up.
Language models calculate which sequence of words is likely to fit. They do not check by themselves whether a statement is true. If a piece of information is missing, the model sometimes fills the gap with something plausible: an opening time, a price, a commitment. On the phone this is delicate, because callers take the information as binding.
The risk can be reduced with several measures: allow answers only from stored knowledge, permit the assistant to say that it does not know something, maintain binding details such as prices and times in structured form, and verify actions such as bookings on the server. Errors cannot be ruled out completely. Spot checks in transcripts and a clear handover to people are therefore part of running such a system.
In practice
In operation you recognise hallucinations by statements that appear in none of your documents. Search the transcripts specifically for numbers, times, prices and promises and compare them with the stored knowledge. If you find an invented detail, the cause is usually a gap: the question came up and the answer was missing. Then add the information, or give the assistant a clear fallback for that subject, such as offering a callback. Repeat the spot check after every larger change to knowledge or instructions.
An example
A caller asks a car repair shop what a brake replacement costs for his model. The stored information contains no price for it. A poorly configured assistant names an amount anyway, and it sounds plausible. The customer arrives at the workshop with that figure, and the quote is considerably higher. Configured properly, the assistant says that the price depends on the vehicle, takes model and registration number and announces a callback with a quote.
A common misunderstanding
Hallucinations are readily treated as a rare slip that a newer model will fix by itself. Newer models err less often, but not never, and they sound just as sure when they do. The opposite mistake is to forbid the assistant any statement that is not stored word for word. Then it no longer answers simple questions either. The sensible split: it may phrase freely, while binding facts come exclusively from maintained sources.
At Call0
Call0 answers questions from approved knowledge. The privacy policy points out that AI outputs can contain mistakes and that important results should be reviewed.
Read on:
On this topic:
As of: 12 October 2026