Practical AI · 3-minute read
My other car gets extraordinary mileage. It has an impeccable safety record, a surprisingly affordable price and one small issue: it exists entirely in a very confident paragraph.
That is the charm of an AI hallucination. The details can be beautifully furnished before anyone notices the building is missing.
In this context, “hallucination” means an AI-generated answer containing false or invented information. It is a technical shorthand, not a claim that the software experiences the world as a person does.
A believable supplier is not necessarily a supplier
Consider Arroyo Lantern Construction, a fictional contractor invented for this example. Its office asks a chatbot for local temporary-fence vendors. In this constructed scenario, the response supplies a business name, a phone number and a neat explanation of next-day availability.
None of those details has been checked. They might look exactly like the information the team needs while being partly or entirely invented.
The office should use the list as research leads. Verify that each business exists through its official site, confirm the contact details, and ask the supplier directly about availability. Do not paste an unverified answer into a purchase order because it arrived wearing bullet points.
OpenAI’s research describes how training and evaluation can reward guessing instead of acknowledging uncertainty. The practical implication is simple: plausible language alone does not establish that a claim is true.
Give the answer somewhere solid to stand
Ask for the basis. Which supplied document or accessible source supports each important claim?
Open the source. A citation can be invented, outdated or irrelevant. Confirm the page exists and says what the answer claims.
Leave blanks visible. “Not confirmed” is more useful than an invented delivery date.
Verify the consequential details. Names, contacts, amounts, dates and commitments deserve direct checks before they drive an action.
Try this prompt
Review this draft for claims needing verification: [draft]. For each important factual claim, identify the supporting passage in the material I supplied or label it “not supported here.” Separate facts, inferences and suggestions. Do not create citations, contact details or missing information. List the questions a person should verify with the original source.
That prompt can help expose gaps; it cannot guarantee that none remain. Keep confidential project details out of an unapproved tool. Check important claims yourself or assign them to a qualified colleague before placing orders or making commitments.
AI is still useful for organizing options, drafting questions and comparing verified information. Just keep imagination in the creative brief and verified facts in the purchase order.
The imaginary car can stay. The imaginary invoice needs to go.
Your turn: What believable detail would be most expensive for your team to get wrong?
Source: OpenAI, “Why language models hallucinate”, September 5, 2025. The contractor and vendor response are constructed examples, not observed outputs or a real client story.
