How to check an AI answer before you use it

You type a question into an AI chat assistant. The answer arrives in seconds, and it looks wonderful. It is tidy, well written and entirely sure of itself. That confidence is the trap.

Many people think that if the AI sounds sure, it must be right. But the AI can sound exactly as sure when it is wrong. It may state a number, a date or a quote that has no source behind it. Sometimes the thing it names does not exist at all.

Think of a fluent speaker at a dinner party. Fluency makes them sound reliable, but it says nothing about whether the facts are right.

The first answer is a draft

None of this is a reason to avoid the tool. It is a reason to read its output with care. The first answer is not the finished work. It is a draft, and you are the editor.

A good editor does not rewrite everything. She reads for what could be false first, and for what could be better second. That gives you two jobs.

Job one: catch what is false

Scan the answer for three kinds of detail. These are the ones that sound most convincing and are easiest to get wrong.

• Numbers: percentages, prices, dates, amounts.
• Names: people, companies, laws, products, places.
• Quotes: anything in quotation marks, or attributed to someone.

For each one, ask where it came from. If you can check it in a minute, check it. If you cannot, ask the AI for its source, then open that source yourself. A source the AI names can be wrong too, so the click matters.

If you cannot find support for a claim, you have three honest options. Drop it. Soften it to what you can support. Or mark it as unchecked and keep it out of anything that goes to another person.

Job two: improve what is weak

For the second job, use a follow-up. A follow-up is a short message that tells the AI exactly what to change in its last answer. Compare two of them.

• “Make it better.” This gives the AI nothing to work with, so it guesses what better means.
• “Cut this to five bullets, one line each. Keep the numbers and mark any claim that needs a source.” This names the length, the shape and what to keep.

Follow-ups also build on each other. The AI still holds the earlier answer, so you do not have to start again or paste it back in.

An edit in practice

Say you ask for a short summary of why customers cancel a subscription. The answer comes back as four long paragraphs. In the second paragraph, one sentence claims that most cancellations happen in the first month. It names no study. That is a claim about most customers, and nothing supports it, so do not copy it into your report.

Send a follow-up: “Which study says this? Give the title and the publisher, or tell me if you cannot find one.” Then treat whatever comes back with care. If the AI cannot name a real source, drop the claim or soften it.

Notice what you did not do. You did not throw the whole answer away, because most of it was probably fine. Now fix the rambling with a second follow-up that asks for five short bullets. The answer should shrink to something you can scan in half a minute.

Where this habit matters most

A lawyer’s assistant asks the AI to summarise a contract clause. The summary sounds right, yet she still checks every date and amount against the contract itself. A restaurant owner asks for a food safety checklist. It includes a temperature, so he checks it against his local authority’s guidance before printing a single copy.

Neither person is being distrustful. They are doing what any editor does with a first draft from a new writer. The rule is the same whether the stakes are small or large: the more it matters if the detail is wrong, the more carefully you check it. And whatever the AI drafts, you are the one who sends it.