# How to prevent AI chatbot hallucinations URL: https://instantknowledgebot.com/learn/ai-chatbot-hallucinations Last updated: 2026-10-14 > AI chatbot hallucinations are fluent but false answers, which happen when a language model fills gaps with plausible guesses. You prevent most of them by grounding the chatbot in your own documents, instructing it to say when the answer isn't there, giving it a fallback message, keeping documents clear and current, and reviewing unanswered questions. **AI chatbot hallucinations** are answers that sound confident and fluent but are wrong: an invented refund policy, a price you never charged, a feature you don't have. For a business chatbot, they're the biggest risk to trust. The good news is that most hallucinations are preventable with the right setup and a little ongoing care. ## What is an AI hallucination? Large language models generate text by predicting likely next words. They're very good at producing plausible language, but plausible isn't the same as true. When a model lacks the information it needs, it may fill the gap with something that *sounds* right. That's a hallucination. Typical examples in business chatbots: - Stating a return window of 60 days when your policy says 30. - Describing a service or product you don't offer. - Giving a phone number or URL that doesn't exist. - Promising delivery dates, discounts or availability. ## Why do chatbots hallucinate? | Cause | Example | Fix | |---|---|---| | **No grounding** | A general AI model answering about your business from its training data | Ground answers in your documents | | **Missing information** | Your documents don't mention weekend hours | Add the information, and set a fallback | | **Contradictory documents** | Two price lists from different years | Remove outdated files | | **Ambiguous questions** | "How much is it?" | Let the bot ask a clarifying question | | **Retrieval misses** | The answer is buried in a huge file under different wording | Split and structure documents | | **Over-eager instructions** | "Always give an answer" | Allow and encourage "I don't know" | ## How do you prevent chatbot hallucinations? **1. Ground the chatbot in your documents.** Use a knowledge base chatbot that gives the model your content for every question and instructs it to answer only from that content. This approach is explained in [RAG chatbots explained](/learn/rag-chatbot-explained). **2. Make "I don't know" an acceptable answer.** The bot should say plainly when the knowledge base doesn't cover something. InstantKnowledgeBot instructs every bot to do this. **3. Set a useful fallback message.** "I don't have that in my notes. Email support@yourcompany.com and we'll help today." A good fallback turns a dead end into a next step. **4. Fix the source documents.** Remove outdated versions, resolve contradictions, put key facts in text (not images), and write specifics rather than vague statements. See [preparing documents for an AI chatbot](/learn/preparing-documents-for-ai-chatbot). **5. Add explicit guardrails for risky topics.** Extra instructions like "Never quote delivery dates" or "Never give medical advice; suggest booking a visit". InstantKnowledgeBot's industry templates include guardrails such as never inventing listings (real estate), never guaranteeing allergen-free dishes (restaurants), and never giving legal or medical advice (law firms and clinics). **6. Test before launch.** Ask real questions, rephrased questions, questions spanning two documents and questions the documents don't answer. Check every number. **7. Monitor after launch.** Review unanswered questions, thumbs-down answers and a sample of conversations each week at first. ## How do you test a chatbot for hallucinations? Build a small test set of 20 to 30 questions: - **Ten common questions** with known answers. - **Five "numbers" questions:** prices, limits, dates, timeframes. - **Five questions not covered** by your documents. The correct answer is "I don't know" plus your fallback. - **Five tricky questions:** different wording, two-document answers, ambiguous phrasing. Run it before launch and after every significant document change. Anything wrong usually points to a document to fix. ## What does a well-grounded answer look like? Say your policy says "Returns accepted within 30 days of delivery for unworn items." A visitor asks: "I bought shoes 6 weeks ago and wore them once. Can I return them?" - **Hallucinated:** "Yes! We offer a 60-day satisfaction guarantee on all footwear." - **Grounded:** "Our returns policy accepts unworn items within 30 days of delivery, so unfortunately these wouldn't qualify. If there's a fault, email support@… and the team can take a look." The grounded answer applies your actual policy, says no politely, and offers a next step. ## Which questions are most likely to trigger hallucinations? Some question types deserve extra testing because they tempt a model to fill gaps: - **"Do you have…?" questions** about products or services you don't offer. The bot should say no, or that it doesn't see it, rather than describe something plausible. - **Questions about the future:** "When will this be back in stock?" or "Will prices go up?" Unless your documents say, the right answer is "I don't know". - **Calculations:** "How much for three people for two nights?" Check the arithmetic and the inputs it used. - **Comparisons with competitors:** keep these out of scope unless you've written factual content about them. - **Personal or account-specific questions:** "Has my order shipped?" A knowledge base chatbot can't know, and should say so. Add one or two of each type to your test set. ## Who should own chatbot accuracy? Make it someone's job. In a small business, that's often the person who already answers customer emails, because they know which answers matter. A light routine is enough: a weekly look at knowledge gaps and thumbs-down answers for the first month, then monthly. Each fix goes into the source document, so the website, the FAQ page and the chatbot all improve together. ## How does InstantKnowledgeBot handle this? - **Grounding:** every bot answers only from its knowledge base. Small knowledge bases are given to the model in full; larger ones are searched for the most relevant passages. - **Honest fallbacks:** the bot says when the answer isn't there, and you can set the exact fallback message. - **Extra instructions** for your own rules. - **Industry templates** with built-in guardrails; browse the [templates](/templates). - **Knowledge gaps:** the bot overview lists questions your documents couldn't answer. - **Feedback and conversations:** thumbs up and down on answers, and full transcripts to review. No tool can promise zero mistakes, and we don't. But with good documents and these safeguards, a knowledge base chatbot is far more reliable than a general AI assistant answering about your business. Learn the basics in [what is a knowledge base chatbot](/learn/what-is-a-knowledge-base-chatbot), see the [customer support chatbot](/solutions/customer-support-chatbot) playbook, or [build one free](/signup). Plan details are on the [pricing page](/pricing). ## Frequently asked questions **Can any chatbot guarantee it will never hallucinate?** No. Grounding and good documents reduce made-up answers dramatically, but no AI system can promise zero errors. That's why fallbacks, testing and monitoring matter. **Why does a chatbot hallucinate even with my documents?** Usually because the answer is missing, outdated or contradictory in the documents, the question is ambiguous, or retrieval didn't find the right passage. Fixing the source content solves most cases. **Does a more expensive AI model hallucinate less?** More capable models often follow grounding instructions better, especially with long or complex documents. But document quality usually matters more than the model. **What should a chatbot say when it doesn't know?** It should say plainly that it doesn't have that information and offer a next step, such as your support email, phone number or a contact form. **How does InstantKnowledgeBot reduce hallucinations?** Every bot is instructed to answer only from its knowledge base and to say when the answer isn't there; you can set a custom fallback message, and the knowledge gaps report shows questions your documents couldn't answer.