To stop a chatbot making things up, restrict it to your own verified content, tell it to say “I don’t know” when the answer is not there, and make it show its sources. Then hand uncertain chats to a person and test it regularly with real customer questions. Clear instructions alone reduce errors, but grounding is what fixes them.
At a glance
- Anthropic’s guidance lists allowing the model to say “I don’t know” as its first basic strategy for minimising hallucinations.
- OpenAI’s guide sets out prompt engineering, retrieval-augmented generation (RAG) and fine-tuning as the main tools for improving accuracy, and covers how much accuracy is “good enough” for production.
- Microsoft’s RAG guidance includes a dedicated section on designing grounding instructions, with separate prompt structures for customer support, legal and technical documentation.
- Chati answers customers from a business’s own website, shows its sources and hands chats to a human team when unsure.
- Chati is trained on the customer’s own data and supports custom integration.
Why do chatbots make things up?
Chatbots make things up, known as hallucination, because language models predict plausible text rather than look up verified facts, so they fill gaps with confident guesses.
Think of it like a new member of staff who hates saying “I’m not sure”. Asked something they have no answer to, they come up with something that sounds right. The chatbot does the same, only faster and with perfect confidence. OpenAI has published its own explanation of how to improve the accuracy of language models, and the starting point is accepting that a model on its own is not a fact checker.
The main causes
- Missing information. If your site never mentions parking, the chatbot may invent a parking policy rather than admit it has nothing.
- Vague questions. “How much does it cost?” could mean a dozen things, and a guess is easier than a follow-up question.
- Outdated data. A price list from last year is still an answer, just a wrong one.
- Rewarded for answering. Models are generally built to be helpful, which can push them towards giving an answer over admitting doubt.
What it looks like in a small business
A made-up opening time on a bank holiday. A quoted price for a service you do not offer. A cancellation policy that sounds reasonable but is not yours. Each one is small, and each one can lead to an annoyed customer at your front desk.
How do you stop a chatbot making things up?
Combine five fixes: ground answers in your own content, set clear rules, allow “I don’t know”, show sources, and add human handover and testing.
- Ground the chatbot in your own website and documents.
- Give it clear written rules about what it may and may not say.
- Let it say “I don’t know” and give it a fallback.
- Show the source page behind every answer.
- Hand uncertain chats to a person, and test and review regularly.
No single fix removes errors entirely. That is why layers matter: if one slips, another catches it. The rest of this article covers each fix in order.
Should you ground the chatbot in your own website content?
Yes. Grounding, or retrieval-augmented generation, makes the chatbot answer only from supplied sources such as your website, which is the most effective fix for invented answers.
RAG sounds technical but the idea is simple. When a customer asks a question, the system first finds the most relevant page or passage from your content. It then writes the answer using only that material. The chatbot is working from your pages, not from memory.
Prompting, RAG and fine-tuning compared
OpenAI’s accuracy guide describes three tools, usually tried in a sensible order.
| Approach | What it means | Best for |
|---|---|---|
| Prompt engineering | Writing clear instructions for the chatbot | Setting tone, rules and the “I don’t know” behaviour |
| Retrieval-augmented generation (RAG) | Fetching your own content and answering from it | Facts that are specific to your business and change over time |
| Fine-tuning | Further training a model on example conversations | Consistent style or format, rather than keeping facts current |
For most small businesses, good instructions plus RAG covers what is needed. Prices and opening hours change, and it is far easier to edit a web page than to retrain a model.
Keep the source content accurate
Grounding only helps if the content is right. If your website says one thing and your price list says another, the chatbot will repeat whichever it finds. Tidy up contradictions, remove old pages and date anything that changes.
This is how Chati works. It answers customers from your own website and is trained on your own data, so the answers come from what you have actually published. You can see the steps on our how it works page.
What instructions should you give a chatbot to stop it guessing?
Tell it to use only the supplied content, never invent facts, figures, prices or sources and separate what is known from what is not.
Here is a block you can adapt and paste into a system prompt, the standing instructions a chatbot reads before every chat:
Answer only using the content provided to you. Do not use outside knowledge. Never invent facts, prices, dates, opening hours, availability, policies or sources. If the content does not clearly answer the question, say you do not know and offer to pass the customer to our team. Do not give legal, medical or financial advice. Keep answers short and say which page your answer comes from.
Rules worth adding
- Prices: quote only prices that appear in the supplied content, and say that they may change.
- Availability and bookings: never promise a slot unless the booking system confirms it.
- Opening hours: use only the published hours and point to the contact page for bank holidays.
- Regulated topics: for anything legal, clinical or financial, give general information from your site and hand over to a person.
Scenario rules for customer support
Microsoft’s RAG guidance has a section on designing grounding instructions, with different prompt structures for customer support, legal and technical documentation. The lesson for you is that one blanket rule is rarely enough. A support question about a delivery needs a different rule from a question about your terms and conditions. Write a short rule for each type of question your customers actually ask.
A warning: instructions alone are not enough. A model told to stick to the facts but given no facts will still be tempted to fill the gap. Instructions work best on top of grounding.
How do you let a chatbot say “I don’t know”?
Give it explicit permission and a fallback, such as offering contact details or passing the chat to a person, so it never feels forced to answer.
This is the first basic strategy in Anthropic’s guidance on reducing hallucinations: explicitly allow the model to admit when it does not know. It sounds almost too simple, but without that permission the chatbot tends to treat every question as one it must answer.
An example fallback message
I’m sorry, I can’t find that on our website, and I don’t want to give you the wrong answer. Feel free to call us or leave your email and we’ll come back to you.
The confidence idea, without the jargon
Behind the scenes, a good setup checks whether the content it found actually answers the question. If the match is weak, the chatbot should not push on. Think of it as a threshold: strong match, answer; weak match, admit it and offer a person. You do not need to set numbers yourself, but ask any provider how their chatbot behaves when the match is weak.
Why should a chatbot show its sources?
Showing the page behind each answer lets customers and staff check it quickly, and makes any error easy to spot and fix at the source.
Another option is to ask the chatbot, in its instructions, to point to the passage it used. If it cannot find one, it should not make the claim. Anthropic’s guidance suggests asking the model to back up claims with quotes from the supplied material, and that works well in plain customer chats too.
Sources build trust
A customer who sees “From our pricing page” feels more confident than one who gets an unexplained number. It also helps your team. When a wrong answer appears in the logs, the source link shows you straight away which page needs correcting.
Chati shows its sources with each answer, so customers can click through and your team can trace where it came from.
When should a chatbot hand over to a human?
It should hand over whenever it is unsure, the question is sensitive, or the customer is ready to book or buy.
Common triggers
- Low confidence, where the content does not clearly answer the question
- Complaints or upset customers
- Disputes about pricing or bills
- Urgent matters
- Regulated or advice-style questions
Examples by sector
- Dental practices: a patient describing pain or asking whether a treatment suits them should reach the team, not get a guess.
- Solicitors: questions about someone’s own legal situation need a person, while general questions about how you work can be answered from your site.
- Letting agents: deposit disputes, repairs and tenancy problems should go to staff.
- Vets: anything that sounds like an emergency should be passed on straight away, with your emergency contact details shown.
Chati can hand chats to a human team when it is unsure. Your team can then pick up the conversation without asking everything again. See our industry guides for how this looks in different trades and professions.
How do you test and monitor a chatbot for wrong answers?
Test it with your 20 to 50 most common real customer questions, check each answer against your site, and review chat logs regularly.
Build a simple test sheet
Open a spreadsheet and list the questions customers really ask, taken from emails, calls and enquiry forms. Add columns for the chatbot’s answer, the correct answer, the source shown and a pass or fail mark. An answer passes only if it is correct and the source backs it up. Keep it simple. Rerun the sheet whenever you change the chatbot or your website.
Try the tricky questions
Questions that often trip chatbots up include:
- Something your business does not offer at all
- A question with a false assumption, such as “Why is your weekend rate cheaper?” when there is no such rate
- Asking for a discount or special deal
- Asking for a guarantee or a legal opinion
- Vague questions like “How much?”
The right result is a clear correction, an “I don’t know” or a handover, not a confident invention.
Review logs and update content
Read your chat logs weekly at first, then monthly once things settle. Look for answers that were wrong or vague, or that ended in a handover, and fix the page behind them. Update your content whenever prices, hours or policies change.
OpenAI’s guide also covers deciding how much accuracy is good enough for production. For a small business, a sensible reading is to set a standard before you go live. Prices and policies might need to be right every time. A chatty question about your team can be looser.
What should you look for when choosing a chatbot?
Choose a chatbot that answers from your own content, shows sources, hands over to people, and lets you review and correct answers.
A short checklist
- Answers from your own website and documents, not general knowledge
- Shows the source behind each answer
- Says “I don’t know” and offers a fallback
- Hands chats to a person when unsure
- Lets you read chat logs and fix the source content
- Is clear about cost, which you can compare on our pricing page
Where Chati fits
Chati is built around this checklist. It answers from your own website, shows its sources and hands chats to a human team when unsure. Branding, lead capture and bookings are included too.
Our team has been developing apps since 2009, and you can read more about who builds Chati on our about page. If you would like a chatbot that sticks to what your business actually says, get in touch with Chati and we will talk it through with you.
Frequently asked questions
Can you stop a chatbot hallucinating completely?
Not completely. Grounding the chatbot in your own content, setting clear rules, showing sources and handing over when unsure will cut it sharply. Regular testing catches what is left.
Does ChatGPT make things up?
Yes, it can. Like other language models, it predicts likely text and can state wrong information confidently, especially when it has no reliable source to work from.
How do I make a chatbot only use my website content?
Use a chatbot that is grounded in your site through retrieval-augmented generation, then add an instruction to answer only from the supplied content and to say when the answer is not there.
What is retrieval-augmented generation?
It is a method where the chatbot first finds the most relevant passage from your own content, then writes its answer from that passage instead of relying on memory.
How often should I check my chatbot’s answers?
Review chat logs weekly when you first launch, then monthly. Check again straight away whenever your prices, opening hours or policies change.
Is a chatbot that gives wrong answers a legal risk for my business?
It can be, particularly if customers rely on wrong prices, terms or advice. This article is not legal advice, so ask your own adviser about your position. Grounding, sources and handover reduce the risk.
Do I need to train a chatbot on my own data?
For accurate answers about your business, yes, in the sense that it needs access to your own content. Our chatbot, Chati, is trained on your data and answers from your website.
