Does your bot never make anything up?

We do not go that far, and you should not believe anyone who does. Our bot answers only from the knowledge provided and says openly when it does not know something; that makes made-up answers rare, but there is no absolute guarantee with language models. This limit is also stated in our terms and conditions.

Are we liable for what the bot says?

Information from your chat window can be attributed to your company; the Air Canada case shows that. All the more important that the bot only says what you have provided, and that you decide through the knowledge what that is.

How do I know what the bot could not answer?

With us, the customer portal keeps the list of questions for which there was nothing in the knowledge. You see the gaps and close them in a few sentences.

Can I stop the bot from talking about certain topics altogether?

Yes, through the knowledge: if something isn’t in it, the bot doesn’t answer it. On prices, for example, you decide whether it gives a range, explains what the price depends on, or refers to a quote.

Guide

When a chatbot makes things up

An AI chatbot always answers fluently and always with full conviction, even when the answer is entirely made up. That is not a rare defect of individual providers but a property of the technology behind it, and anyone who puts a bot on their website should understand it.

This page explains why it happens, what it can cost, and how you can tell before you buy how a provider deals with it. Including us.

Why an AI makes things up at all

A language model is built to produce a fluent, plausible-sounding text for every question. It has no built-in notion of ‘I don't know’: where knowledge is missing, it fills the gap with whatever fits best linguistically. The result sounds just as certain as a correct answer, and that is exactly what makes it dangerous.

For a chatbot on a company website, this means: without being tied firmly to the business's knowledge, it will also answer questions about prices, deadlines and commitments it knows nothing about. Not out of malice, but because filling gaps is in its nature. The whole craft lies in forbidding it to do exactly that.

What a made-up answer costs

The damage comes in two sizes. The small one: a wrong piece of information sticks with the customer, and you clear it up in conversation, each time at the price of a little trust. The large one: you are held to what was said.

The case	What happened
Air Canada, 2024	The airline's chatbot gave a customer wrong information about a refund rule. A Canadian tribunal ruled that the airline is just as responsible for what its bot says as for what an employee says. Air Canada paid the difference.

Moffatt v. Air Canada, Civil Resolution Tribunal (Canada), decision of February 2024.

The decision comes from Canada, but the idea behind it is an obvious one elsewhere too: what your chat window promises a customer, your company promises. A bot that invents commitments invents them in your name.

What helps against it

Tying it to the knowledge provided: The most effective protection: the bot may only answer from what the business has provided, not from the language model's general knowledge. That is how our bot is built. If there is no delivery time in the knowledge, there is nothing the bot could turn into one. It then says that it does not know.
Admitting the gap: If the bot finds nothing in the knowledge for a question, it says so instead of guessing, and takes down the request with contact details. An admitted gap costs a moment; a made-up answer costs trust, or more.
The route to a person: If a visitor asks for a person, is annoyed, reports something that has gone wrong, or wants a binding commitment, the bot passes the conversation to the business, tells the visitor it has done so, and files it in the customer portal. That applies on every plan.
The list of gaps: Every unanswered question appears in the customer portal. The business sees what was missing and adds to the knowledge. That way every gap becomes a better answer for the next visitor.

The test before you buy: three questions for any bot

You can check any provider in five minutes, on their own demo bot. Ask three questions:

A question whose answer is certainly in the knowledge, such as the price. This checks the basic quality.
A question from the subject area whose answer is certainly NOT in the knowledge, such as one about a made-up product or a very specific exception. A sound bot says it does not know. A dangerous one answers anyway, fluently and wrongly.
A question entirely off topic, such as the weather or a recipe. This shows whether the bot sticks to its job or turns into a chatterbox whose answers you have to answer for.

You are welcome to run this test on us too, on our sample website with the bot (in German). That is exactly what it is there for.

Common questions about made-up answers

Does your bot never make anything up?: We do not go that far, and you should not believe anyone who does. Our bot answers only from the knowledge provided and says openly when it does not know something; that makes made-up answers rare, but there is no absolute guarantee with language models. This limit is also stated in our terms and conditions.
Are we liable for what the bot says?: Information from your chat window can be attributed to your company; the Air Canada case shows that. All the more important that the bot only says what you have provided, and that you decide through the knowledge what that is.
How do I know what the bot could not answer?: With us, the customer portal keeps the list of questions for which there was nothing in the knowledge. You see the gaps and close them in a few sentences.
Can I stop the bot from talking about certain topics altogether?: Yes, through the knowledge: if something isn’t in it, the bot doesn’t answer it. On prices, for example, you decide whether it gives a range, explains what the price depends on, or refers to a quote.

Run the three-question test on us.

Try our bot on the sample website and ask it something it cannot know. And if you want to talk afterwards: we reply within 24 hours, in person.

Try the bot (in German)

Last checked: 4 October 2026