Glossary
AI hallucination
In short
An AI hallucination is when an AI model states something false or invented as if it were fact, such as a price you don't charge, a service you don't offer or a policy that doesn't exist.
Hallucinations happen because language models generate plausible-sounding text rather than looking facts up. When the information they were given has no answer, they may fill the gap with a guess.
For a business agent this matters more than almost anything else. The usual safeguards: answer only from an approved knowledge base, say "I don't know" and offer a person instead of guessing, limit what the agent may promise, and review transcripts regularly.
No setup removes the risk completely, so anything with real consequences (refunds, legal or medical questions, custom quotes) should go to a person.
A worked example
In 2024 a Canadian tribunal ordered Air Canada to compensate a customer after the airline's website chatbot described a refund policy that did not exist. The airline argued the chatbot was responsible for its own words; the tribunal disagreed. For a small business the lesson is the same: whatever your agent says, customers will hold you to it.
When it is worth worrying about, and when it isn't
Worry about it wherever an answer carries money, safety or legal weight: prices, refunds, medical or legal questions, promises about dates. It matters far less for low-stakes answers such as opening hours or directions, as long as those come from your own information.
Common questions
Can hallucinations be stopped completely?
No. They can be made rare by answering only from approved information and by telling the agent to say "I don't know" rather than guess, but no setup removes them entirely.
How do I spot them?
Read a sample of transcripts every week, and look especially at answers about prices, policies and promises.
Are newer AI models better at this?
Generally yes, but the bigger difference comes from the setup: what the agent is allowed to answer, what it answers from, and when it hands over.