Why your AI chatbot
makes things up.
Your chatbot quoted a price you never set, opening hours you don't keep, a refund policy you've never written. It wasn't broken — it was working exactly as designed. From the HyperChat blog: why that happens, and the three layers that stop it.
Chatbots make things up because language models predict plausible-sounding text rather than looking facts up — fluency is the product, truth is optional. The fix is three layers: ground the bot in your actual website content, exam it before launch, and gate it at runtime so it says "I don't know" and captures the visitor's details instead of guessing.
It's not lying.
It's autocomplete with confidence.
A large language model has one job: produce the most plausible next words. Nothing in that job description says "check whether they're true".
Asked your prices, the model doesn't search your site — it generates what an answer to that question usually sounds like. If your real price list isn't in front of it at that moment, the plausible figure wins over the true one.
The made-up answer arrives in the same polished, helpful tone as the correct ones. There's no hesitation, no asterisk, no "I think". Visitors can't tell the difference — which is precisely the problem.
The model has no built-in sense of "I don't know this". Left to itself it will always answer, because answering is what it does. Uncertainty has to be engineered in — it never arrives by default.
This is why "hallucination" isn't a bug one vendor can patch. It's the default behaviour of the underlying technology — every chatbot inherits it, and the honest ones engineer around it.
A wrong answer costs more
than a missed one.
A missed enquiry is worth nothing. A wrong answer can cost less than nothing — it arrives with a screenshot attached.
The bot quotes a figure it found nowhere on your site. Now you're honouring a price you never set, or starting the relationship by telling a customer their screenshot is wrong. Neither option is free.
"Yes, open Bank Holiday Monday." You're not. The visitor makes the trip, finds the shutters down, and tells Google about it. One hallucination becomes a one-star review that outlives the chat.
"Of course, full refunds any time." Companies have already been held to promises their chatbots fabricated. For a small business, the refund, the complaint and the lost trust all land on the same desk — yours.
Do your own maths: how many chats does your site get a week, and what's one customer worth to you? If even a small share of those conversations ends in a confident wrong answer, the cost of "the bot guessed" quickly dwarfs the cost of fixing it.
Three layers between the bot
and a confident lie.
You can't prompt hallucination away — "never invent facts!" is an instruction, not enforcement. What works is engineering, in three layers:
1 · Grounded retrieval
Before the bot answers, it looks up the relevant passages from your actual website and answers only from them. The model still writes the sentence — but the facts come from your pages, not from its imagination.
2 · A pre-flight exam
After every knowledge-base build, the bot answers ~10 questions drawn from your own site, scored for groundedness, plus five adversarial probes — injection, jailbreak, price-fishing, off-topic, policy-fishing. You see the score before any visitor does.
3 · Runtime guardrails
Every message is screened for injection before it costs a penny, and when confidence is low the bot says "I don't know" instead of guessing — then offers to take the visitor's details. A dead end becomes a captured lead.
Want the full mechanics — how the exam is marked, what the probes fire, how the honesty gate decides? Read the deep dive: the AI chatbot that doesn't make things up.
Hallucination, asked
plainly.
Why does my chatbot give confident answers that turn out to be wrong?
Is making things up a bug I can patch out?
Can I fix hallucinations just by writing a better prompt?
What should a chatbot do when it genuinely doesn’t know?
How can I tell if my current chatbot is making things up?
Does HyperChat’s anti-hallucination system cost extra?
Keep reading: the accuracy deep dive · more from the blog · pricing
Test a bot that admits "I don't know".
Build it free from your website in 60 seconds, then try to make it guess. It won't. No card, no commitment.
Build my chatbot · free