AI glossary for businesses
AI hallucination
In short
An AI hallucination is an answer from an AI model that sounds confident and plausible but is false, unsupported or invented. Examples include made-up facts, wrong figures, fake quotes or sources that do not exist. It happens because language models generate likely-sounding text rather than checking facts against a trusted source.
How hallucinations happen
A large language model writes by predicting what text is likely to come next. Most of the time, likely text is also true text. But when the model lacks the right information, it does not stop. It fills the gap with something that sounds right.
Think of a confident student in an oral exam who did not study one chapter. Rather than say "I don't know", they give a fluent answer with invented details. The tone gives no warning. That is what makes hallucinations risky: a wrong answer looks exactly like a right one.
Hallucinations are more likely when you ask about niche or very recent topics, exact numbers, dates, legal or tax details, or sources and links.
In practice for a small business
A small online shop adds an AI chatbot to answer customer questions. A customer asks whether opened cosmetics can be returned. The policy was never given to the bot, so it replies, politely and clearly, that returns are accepted within 30 days for any reason. The shop's real policy says otherwise. Now there is an unhappy customer and a possible dispute.
The fix is not to drop AI. It is to change the setup:
give the tool your actual policies and tell it to answer only from them, an approach called RAG;
instruct it to say "I'm not sure, let me pass you to a colleague" when the answer is not in the documents;
review a sample of answers every week;
keep a human check for anything involving money, health or legal terms.
Why it matters and what to watch out for
Responsibility stays with you: customers will treat what your chatbot or AI-drafted email says as your company's word.
Do not trust AI on legal or tax facts: verify against official sources and check with your accountant, DPO or legal advisor.
Check sources: if a tool gives links or references, open them. Some may not exist.
Personal data: a hallucination about a real person can create GDPR accuracy issues. Be careful with outputs that describe customers or staff.
It can be reduced, not removed: newer models make fewer errors, but none are error-free. Plan your review steps accordingly.