Ask a chatbot something it does not know, and sometimes it will not say "I don't know". It will give a confident, fluent, wrong answer. This is called a hallucination, and for a business it is one of the biggest risks of using AI with customers.
You cannot remove the risk completely, but you can reduce it dramatically. This guide explains why hallucinations happen and gives eight controls that work in real business chatbots.
Why do chatbots hallucinate?
Language models are trained to produce likely-sounding text, not to check facts. When the model lacks the information, or the question is unclear, it may still produce a plausible answer, because plausible text is exactly what it is good at.
Hallucinations get more likely when:
- The question needs information the model never saw (your prices, policies, or recent changes)
- The instructions do not tell the bot what to do when it does not know
- The question is vague, or the bot is allowed to talk about anything
- Retrieved source material is missing, outdated, or contradictory
Why it matters for your business
A wrong answer about a refund policy, a price, or a delivery time can cost money and trust. In the well-known case covered in our article on chatbot liability, a company was held responsible for its chatbot's wrong statement. The goal is not a perfect model. The goal is a system designed so that wrong answers are rare, caught, and not harmful.
Control 1: Ground answers in your approved content
The most effective single step. Instead of letting the model answer from memory, retrieve the relevant passages from your own documents and instruct the model to answer only from them. This is the idea behind RAG. See our guides to RAG chatbots and RAG vs fine-tuning.
Control 2: Tell the bot what to do when it does not know
Write explicit instructions:
Answer only using the information provided. If the answer is not in the information, say you are not sure and offer to connect the customer with a team member. Never guess prices, dates, or policies.
Our guide to chatbot prompt design shows how to structure rules like these.
Control 3: Keep the knowledge base clean and current
Outdated or contradictory documents produce wrong answers even with perfect retrieval. Assign an owner, set review dates, remove duplicates, and update content when products or policies change.
Control 4: Limit the scope
A bot that can discuss anything will eventually say something wrong about something. Limit it to your business topics and have it politely decline the rest. Narrow scope also helps with platform rules, as explained in our article on WhatsApp's chatbot policy.
Control 5: Show sources
When the bot answers, show where the answer came from, such as a link to the help article. This lets customers verify, helps your team spot errors, and builds trust.
Control 6: Handle high-risk topics differently
Some topics deserve special handling:
- Prices, refunds, contracts, legal, medical, and financial questions: use fixed approved answers, link to the official page, or hand over to a person.
- Actions with consequences: require confirmation or approval.
Control 7: Validate before you send
For important outputs, add checks:
- Compare with a source of truth. Look up the actual price or order status from your system instead of letting the model recall it.
- Structured outputs. Ask for answers in a fixed format that can be checked automatically.
- A second check. A second model or rule set reviews the draft answer against the source.
These steps add cost and delay, so use them where mistakes matter most.
Control 8: Test, monitor, and keep improving
You cannot control what you do not measure.
- Test before launch with real and tricky questions. Our chatbot testing checklist covers what to include.
- Sample real conversations every week and score accuracy.
- Track unanswered and escalated questions. They show content gaps.
- Collect customer feedback (a thumbs up or down on each answer).
- Fix quickly and add each mistake to your test set so it does not return. See our guide to chatbot KPIs.
What does not work well
- Telling the model "do not hallucinate" and nothing else. It helps a little, but it is not a control on its own.
- Choosing a bigger model and expecting the problem to vanish. Better models hallucinate less, but they do not stop.
- Skipping human handover. Always give customers a way to reach a person.
- Ignoring your content. The best model with bad documents still gives bad answers.
Be honest with customers
State that the customer is talking to an automated assistant, and encourage them to confirm important details with official sources. A disclaimer does not replace accuracy, but transparency is part of responsible design.
A quick checklist
- Answers grounded in approved content
- Clear "I'm not sure" behaviour
- Clean, current knowledge base
- Defined scope
- Sources shown to users
- Special handling for high-risk topics
- Validation for critical outputs
- Testing before launch and monitoring after
Types of chatbot mistakes (and how to tell them apart)
Not every wrong answer is the same kind of failure. Knowing the type tells you the fix.
| Type | Example | Usual cause | Usual fix |
|---|---|---|---|
| Invented fact | States a policy that does not exist | Model guessed without a source | Grounding, "do not guess" rules |
| Outdated fact | Quotes last year's price | Old documents in the knowledge base | Content updates, re-indexing |
| Wrong source | Answers about product A using product B's page | Poor retrieval or missing metadata | Better chunking, filters, re-ranking |
| Misread question | Answers a different question | Vague wording, no clarifying step | Clarifying questions, better instructions |
| Overconfident tone | Sounds sure when evidence is weak | No uncertainty rules | "I'm not sure" behaviour, thresholds |
| Out-of-scope answer | Gives advice outside your business | Scope not limited | Scope rules, refusal examples |
| Action error | Books the wrong slot | Faulty tool use or missing validation | Confirmation steps, validation |
When you log a mistake, label its type. Over a few weeks, patterns appear and show where to invest.
How to measure your hallucination rate
You cannot reduce what you do not measure. A simple method:
- Sample. Each week, pick 50 random conversations (more for high-risk topics).
- Score each answer as correct, partly correct, or wrong, using your approved content as the source of truth. Have two people score a subset to check consistency.
- Calculate the rate of wrong answers, and separately the rate of wrong answers on high-risk topics.
- Label the cause using the table above.
- Fix and re-test. Add each failure to your test set. Our chatbot testing checklist shows how.
- Track the trend over time alongside your other chatbot KPIs.
Set a target by risk level. For routine questions you might tolerate occasional minor errors. For prices, refunds, and legal or medical topics the tolerance should be close to zero, which is why those topics need special handling.
An incident response plan
Even careful teams will see a serious wrong answer sooner or later. Decide in advance what to do:
- Detect. Customer complaint, internal review, or monitoring alert.
- Contain. Fix the answer immediately, for example by correcting the source content or adding a rule. If needed, disable the topic or route it to humans.
- Assess impact. Who saw it, and did anyone act on it? Do you need to contact customers?
- Correct. Reach out to affected customers if appropriate, and put things right.
- Learn. Find the root cause, update tests, and improve the system.
- Record. Keep a log of what happened and what you changed. In a dispute, records of reasonable care matter. See the lessons from the Air Canada chatbot case.
Frequently asked questions
Can AI hallucinations be eliminated completely?
No. They can be reduced a lot with grounding, clear scope, validation, and monitoring, but you should design for the possibility of mistakes and make them rare, visible, and harmless.
Which is better at avoiding hallucinations, a bigger model or RAG?
Both help, but grounding answers in your own content (RAG) is usually the bigger improvement for business facts. See RAG vs fine-tuning.
How often should I review chatbot answers?
Weekly in the first months, then at least monthly, plus whenever you change content, prompts, or models.
Should the chatbot show a disclaimer?
It is good practice to say it is an automated assistant and that important details should be confirmed with official sources. A disclaimer does not replace accuracy.
Do hallucinations happen in voice and WhatsApp bots too?
Yes. The same controls apply. For voice, also plan confirmation steps for names and numbers. See our guide to AI voice agents.
Build a chatbot that stays on the facts
We design chatbots with grounded answers, clear scope rules, testing, and human handover. See our chatbot development service, or book a free consultation.
