An AI hallucination is an answer from an AI model that sounds confident but is false or not supported by any source. It might be an invented price, a policy that does not exist, a fake citation or a wrong date. The model is not lying on purpose: it produces likely-sounding text, and likely is not the same as true.
Why it matters for your business
A hallucination in a private draft is a nuisance. A hallucination sent to a customer is a promise you may have to keep. In Moffatt v. Air Canada, a British Columbia tribunal held the airline responsible for wrong fare information from its website chatbot in February 2024. Whatever the law where you are, customers will hold you to what your bot said.
It also helps to know why it happens. OpenAI researchers argued in a September 2025 paper that common training and evaluation methods reward a model for guessing rather than admitting uncertainty, much like a student guessing on a multiple-choice test. So the problem is built into how these models are scored, not a bug one update will fix.
An example
A customer asks a furniture shop's chatbot, "Do you deliver to Sharjah, and is it free?" The shop's documents cover Dubai delivery only. A poorly set-up bot replies, "Yes, free delivery to Sharjah within 3 days." Nothing in the shop's material says that. The model filled the gap with the most natural-sounding answer. A well-set-up bot says, "I am not sure about Sharjah; let me get a team member to confirm," and alerts the staff.
What to do about it
You cannot remove hallucinations completely. Anthropic's own documentation says its techniques "significantly reduce hallucinations, they don't eliminate them entirely". What you can do is make them rarer and catch them before they cost you:
| Method | What it does |
|---|---|
| Allow "I don't know" | Gives the model an honest exit instead of a guess |
| Answer only from your documents (RAG) | Ties replies to text you approved |
| Ask for quotes or sources | Makes each claim checkable |
| Fixed paths for prices, refunds and bookings | Takes the highest-risk answers away from free text |
| Human review and logs | Catches what slips through |
The first three come from Anthropic's guidance. The last two are our own practice. For the plain explanation of how LLMs work and why they guess, read What is an LLM?. Building these guardrails in from day one is part of our custom AI agents work.
Frequently asked questions
Do newer AI models still hallucinate?
Yes. Newer models are generally better, but every major provider still documents the problem. Treat any model as able to be wrong, and design the system so a wrong answer is caught or harmless.
How do I know if my chatbot is hallucinating?
Test it with questions your documents do not answer and see whether it admits that. Then review real conversation logs every week, looking for prices, dates or promises that do not match your material.
Is a hallucination the same as a mistake in my data?
No. If your price list is wrong and the AI repeats it, that is a data problem. A hallucination is when the AI states something that is not in its sources at all. Both need fixing, in different places.
Last reviewed 22 Oct 2026 by the Astrafolio team.