AI doesn't look facts up. It predicts what a plausible answer sounds like, and most of the time plausible happens to be true. Sometimes it isn't, and the AI delivers the invention in exactly the same confident tone. That gap has a name: hallucination.
What hallucinations look like in practice
They are rarely wild claims. They are boringly specific: a form number that doesn't exist, a deadline moved to a date that was never real, a "small business exemption" you would love to qualify for. The specificity is what makes them convincing. Real damage follows: missed filings, wrong prices quoted to customers, confident answers that turn into refunds.
The three things to always verify
You don't need to fact-check everything. Check anything that is a number, a name, or a rule:
- Numbers: deadlines, fees, percentages, phone numbers.
- Names: forms, laws, agencies, programs. If an AI names something, confirm it exists.
- Rules: especially bright-line rules ("anyone over 20 hours must be W-2"). Clean shortcuts are the most commonly hallucinated thing, because they are what everyone hopes exists.
A habit that takes one minute
Before an AI answer touches a customer, an invoice, or a government form, find one primary source for each specific claim: the agency page, the actual statute, your own records. If you can't find the form or law by name, the AI probably invented it.
Where this bites hardest
Anywhere AI output flows straight to the outside world: auto-drafted quotes, AI chat on your website, generated FAQ pages. Those all deserve a human checkpoint or tight guardrails. Mapping those checkpoints is most of what an AI audit is.
Want to test your eye first? Play Hallucination Hunter: eight confident AI answers, each hiding one invention. Find them before a customer does.