Hallucination, and how a business owner catches one in a costing sheet

A hallucination is a confident wrong answer. The format is right, the arithmetic ties, and one input inside it was never true. This is the normal failure mode of ChatGPT, not a rare glitch. For an owner the danger is never obvious nonsense. It is one plausible number sitting in a sheet you are about to sign off.

What does a hallucination look like in a costing sheet?

A hallucination in a costing sheet looks like a finished sheet. Say you are comparing landed cost for the same consignment into three export markets. You describe the shipment, the volume and the three ports, then ask for a comparison table. The answer comes back tidy. Freight, insurance, duty, clearing, inland movement, total cost per ton, one column per market.

Every column adds up. The arithmetic inside the table is consistent, which is exactly why the sheet reads as already checked. One line in it was produced rather than looked up. The duty rate against one market is a clean figure that no tariff schedule and no clearing agent ever supplied.

Nothing in the output marks that line as weaker than the others. A wrong duty rate does not announce itself in red. It moves the total cost per ton, it moves the ranking of the three markets, and you take a sourcing decision on the ranking. The weak point is one cell whose source you never asked for.

How do I use ChatGPT for work and still trust the numbers?

You check three claims before the output goes anywhere. Not every claim, three. The line in my 30-day guide puts it plainly: "Stress-test the output. Fact-check 3 claims the AI made. Note where it was accurate, vague, or wrong. This is your evaluation muscle."

Pick the three by money, not by doubt. In a landed-cost sheet the three are the duty rate, the freight rate and the exchange rate. Those three decide the ranking. Check each against the source you would have used anyway: the tariff schedule, a quote from your forwarder, the bank rate on the day.

Then write next to each one what you found. Do that ten times and you build a feel for where this output goes soft. Mine goes soft on rates, on dates, and on anything local and recent.

The check takes a few minutes. The sheet took less time than that to produce. That ratio is the whole argument for running the check every time, including on the sheets that look right.

What else do owners ask about wrong AI answers?

How do I learn AI if I cannot spot the mistakes yet?

Start with work whose answer you already know. Run last quarter's numbers, a market you sell in, a supplier you have dealt with for years. Your own knowledge is the answer key. You cannot learn to catch a wrong output on a subject where you have no view.

Does a paid plan stop the wrong numbers?

No. Paying more does not change whether a figure was looked up. Web search helps on anything public and does nothing for a rate in a circular nobody indexed. The three-claim check stays yours on every plan.

Can I ask ChatGPT to check its own work?

You can ask, and the second pass often repeats the same wrong figure in firmer wording. Asking for sources is more useful, because you then check the source yourself. A claim with no source you can open is a claim to verify offline.