How it works
How to stop a chatbot from making things up
Hallucinated prices and invented policies destroy trust. The four controls that keep a website chatbot honest, and how to test for them.
In short
A chatbot makes things up when it is asked something your content does not cover and has not been told to refuse. The fix is grounding plus an explicit instruction to admit uncertainty and hand off.
Test it deliberately: ask questions your site cannot answer and confirm the bot declines rather than improvises.
Key facts
Key facts
- Hallucination risk rises when the knowledge base is thin.
- An explicit 'say you don't know' instruction dramatically reduces invention.
- Support-mode bots should refuse anything outside verified content.
- Test with deliberately unanswerable questions before launch.
The four controls
- · Ground every answer in indexed content from your own site.
- · Instruct the bot to answer only from that content.
- · Give it a refusal path: admit uncertainty, then offer a human.
- · Fill the content gaps the conversations reveal.
The test that matters
Ask five questions your website genuinely does not answer — a price for a service you do not offer, a delivery time to another country, a policy you have never written down. A safe bot says it does not have that information and offers to pass you to a person. A dangerous one invents a plausible number.
Thin content is the root cause
Most invented answers trace back to a website that never explained something. Every refusal in your conversation log is a page you should write, which is why reading the logs improves the site as well as the bot.
The product behind this guide
A chatbot trained on your own website, live today
Defaltam builds AI chatbots that read your public pages and answer with your real prices, services and policies. Support is €79/month, Sales is €149/month, both start with a 5-day free trial, and installation is one script tag.
FAQ
Questions people ask about this
Why do chatbots make things up?
Because a language model will produce a plausible answer when it has no facts and no instruction to refuse. Grounding answers in your indexed content and explicitly telling the bot to admit uncertainty removes most of the risk.
How do I test for hallucinations?
Ask five questions your site does not answer and check the bot declines and offers a handoff instead of inventing details.
Can it be eliminated completely?
Not absolutely, but with grounding, a refusal instruction and good content it becomes rare, and the remaining cases show up in your conversation log where you can fix them.
Does Defaltam's support mode refuse unknown questions?
Yes. Support mode answers only from verified site content and hands off when it is unsure.
Related
Keep reading.
How AI chatbots actually work →
From the visitor's question to the answer on screen: retrieval, context, the language model and why grounding in your content matters.
Running a multilingual AI chatbot →
How a modern chatbot answers in the visitor's language from a single-language website, where it goes wrong, and what to translate anyway.
Website chatbots and GDPR: what you need in place →
What data a website chatbot collects, what your privacy policy must say, and how to handle transcripts and consent under GDPR.
Designing a chatbot handoff to a human →
When a bot should stop answering, what it should collect before passing over, and how to make the handoff feel like service rather than failure.
Website chatbot security basics →
Prompt injection, data leakage, abuse and rate limits — the security questions to ask before putting an AI chatbot on a public website.
Why most website chatbots fail →
Chatbots fail for four predictable reasons: no clear job, thin content, no handoff and nobody reading the logs. Each has a straightforward fix.
AI chatbot plans and pricing →
Support €79/mo, Sales €149/mo, 5-day free trial.
How to add an AI chatbot to your website →
The ten-minute install guide, platform by platform.